Brain-like visual neural network with forward learning and meta-learning capabilities
By introducing explicit position coding and bidirectional neural pathways into brain-like visual neural networks, combined with bioplasticity mechanisms, the problem of high cost of position coding and training of existing deep learning visual algorithms is solved, and rapid visual feature abstraction and generalization are achieved Improve capability.
Patent Information
- Application Number
- CN202010424999.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-05-19
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2040-05-19
AI Technical Summary
Existing deep learning vision algorithms lack explicit encoding of visual feature locations, difficult to describe shapes and contours, rely on error backward transmission and gradient descent training, lack of reverse neural pathways and multi-dimensional information combination mechanisms, resulting in weak recognition generalization ability and training. High cost.
Design a brain-like visual neural network with forward learning and meta-learning functions, adopts implicit and explicit position encoding, including forward and reverse neural pathways, and uses biorational plasticity mechanisms, including multiple learning paradigms, Breakthrough of the von Neumann architecture.
It realizes rapid encoding and abstract visual features, improves recognition and generalization capabilities, reduces training data and cycles, supports multi-dimensional information processing, does not rely on error back-transmission and gradient descent, and is suitable for neuromimicry chip deployment.
Smart Images

Figure CN113688980B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of brain-like visual algorithms and pulse neural networks, and more specifically to a brain-like visual neural network with forward learning and meta-learning functions. Background Art
[0002] Existing deep learning vision algorithms have the following problems:
[0003] 1. Lack of explicit coding of the position of visual features makes it difficult to flexibly describe the positional relationship between features, which is not conducive to encoding and identifying rich and accurate shapes and contours, as well as describing the positional relationship between objects;
[0004] 2. The end-to-end training paradigm that relies on error backpropagation and gradient descent involves a large number of partial differential operations, which has a high training cost and is difficult to break through the von Neumann architecture;
[0005] 3. There is a lack of mechanisms to combine and abstract multiple information from multiple dimensions, making it difficult to form information channels with multiple information dimensions and levels of abstraction;
[0006] 4. It only has forward neural pathways but lacks reverse neural pathways, which cannot support top-down information processing;
[0007] 5. It does not have the forward learning function, and it is not easy to quickly remember the input images or video streams that have been seen, which also leads to a large amount of data required for training and a long training cycle.
[0008] The visual nervous system of the biological brain provides an excellent reference blueprint for designing brain-like visual neural networks.
[0009] According to the neural circuits and working principles of the biological visual nervous system, the brain-like visual neural network should include at least two position encoding methods: implicit position encoding and explicit position encoding. The implicit position encoding is to make the neurons of each layer of encoding features have corresponding receptive fields by correspondingly connecting from the picture to each layer of neurons step by step, instead of using a special neural circuit to encode the position information. This method is not flexible enough, and cannot flexibly combine various visual features at any position, nor can it combine, abstract, and process information for position information, and the generalization ability of recognition is weak. The explicit position encoding is to use a special neural circuit to encode the position information, which can flexibly combine various visual features at any position, and can also combine, abstract, and process information for position information. It can encode richer shape and position relationships, has a strong generalization ability for recognition, and can also accurately recognize situations with strong shape and position relationship constraints.
[0010] The biological visual nervous system also has a bidirectional neural pathway from bottom to top and from top to bottom, which has a priming effect and can help the visual search process. Brain-like visual neural networks should also learn from this feature.
[0011] The biological visual neural system is centered on plasticity mechanisms and has multiple learning paradigms such as reinforcement learning, forward learning, and meta-learning. If the brain-like visual neural network adopts a biologically plausible plasticity mechanism, it can get rid of the training paradigm of error back propagation and gradient descent, avoid a large number of partial differential operations, and is expected to break through the von Neumann architecture, making it more suitable for deployment in firmware or neuromorphic chips; in addition, the brain-like visual neural network should also have forward learning and meta-learning functions, quickly learn and encode the visual features of the pictures or video streams seen, perform information abstraction, find common representations between objects, make generalization ability better, reduce the data required for training, and shorten the training cycle. Summary of the invention
[0012] In view of this, the present invention provides a brain-like visual neural network with forward learning and meta-learning functions, thereby effectively solving the above problems.
[0013] In order to achieve the above object, the present invention adopts the following technical solution:
[0014] A brain-like visual neural network with forward learning and meta-learning functions, comprising: a plurality of primary feature encoding modules, a plurality of composite feature encoding modules;
[0015] Each module includes multiple neurons;
[0016] The neurons include primary feature encoding neurons, concrete feature encoding neurons, and abstract feature encoding neurons;
[0017] The primary feature encoding module includes a plurality of primary feature encoding neurons, encoding primary visual feature information;
[0018] The composite feature coding module includes a concrete feature coding unit and an abstract feature coding unit;
[0019] The concrete feature encoding unit includes a plurality of the concrete feature encoding neurons, encoding concrete visual feature information;
[0020] The abstract feature encoding unit includes a plurality of abstract feature encoding neurons, encoding abstract visual feature information;
[0021] In the expression, if a unidirectional connection is formed between neuron A and neuron B, it means a unidirectional connection of A->B; if a bidirectional connection is formed between neuron A and neuron B, it means a bidirectional connection of A<->B (or A->B and A<-B);
[0022] If there is a unidirectional connection between neuron A and neuron B, A is called the direct upstream neuron of neuron B, and neuron B is called the direct downstream neuron of neuron A; if there is a bidirectional connection between neuron A and neuron B, A and neuron B are called the direct upstream neuron and the direct downstream neuron of each other;
[0023] If there is no connection between neuron A and neuron B, but there are several other neurons connecting them, such as A->C->…->D->B, then neuron A is called the indirect upstream neuron of neuron B, neuron B is called the indirect downstream neuron of neuron A, and neuron D is called the direct upstream neuron of neuron B;
[0024] The excitatory connection is: when the upstream neuron of the excitatory connection emits, a non-negative input is provided to the downstream neuron through the excitatory connection;
[0025] The inhibitory connection is: when the upstream neuron of the inhibitory connection emits, a non-positive input is provided to the downstream neuron through the inhibitory connection;
[0026] Several of the primary feature encoding neurons respectively form unidirectional or bidirectional excitatory / inhibitory connections with other several of the primary feature encoding neurons;
[0027] A plurality of the primary feature coding neurons respectively form unidirectional or bidirectional excitatory / inhibitory connections with a plurality of the concrete feature coding neurons or a plurality of the abstract feature coding neurons located in at least one of the composite feature coding modules;
[0028] A plurality of the concrete feature coding neurons located in the same composite feature coding module respectively form unidirectional or bidirectional excitatory / inhibitory connections with a plurality of the abstract feature coding neurons located in the same composite feature coding module;
[0029] Several of the concrete feature coding neurons and the abstract feature coding neurons in several of the composite feature coding modules respectively form unidirectional or bidirectional excitatory / inhibitory connections with several of the concrete feature coding neurons and the abstract feature coding neurons in other several of the composite feature coding modules;
[0030] The neural network caches and encodes information through the firing of the neurons, and encodes, stores, and transmits information through the connections between the neurons;
[0031] Input a picture or a video stream, multiply a plurality of pixel values of a plurality of pixels of each frame of the picture by weights and input them into a plurality of the primary feature coding neurons, so that the plurality of the primary feature coding neurons are activated;
[0032] For a number of the neurons, their membrane potentials are calculated to determine whether to fire, and if so, the membrane potentials of the downstream neurons are accumulated to determine whether to fire, so that the firing is propagated in the neural network; the weight of the connection between the upstream neuron and the downstream neuron is a constant value, or is dynamically adjusted through the synaptic plasticity process;
[0033] The working process of the neural network includes: forward memory process, memory triggering process, information aggregation process, directional information aggregation process, information transcription process, memory forgetting process, memory self-consolidation process, information component adjustment process, reinforcement learning process, novelty signal modulation process, and supervised learning process;
[0034] The synaptic plasticity process includes a monopolar upstream release-dependent synaptic plasticity process, a monopolar downstream release-dependent synaptic plasticity process, a monopolar upstream and downstream release-dependent synaptic plasticity process, a monopolar upstream pulse-dependent synaptic plasticity process, a monopolar downstream pulse-dependent synaptic plasticity process, a monopolar pulse time-dependent synaptic plasticity process, an asymmetric bipolar pulse time-dependent synaptic plasticity process, and a symmetric bipolar pulse time-dependent synaptic plasticity process;
[0035] Mapping a number of the neurons to corresponding labels as output.
[0036] In one embodiment of the present invention, several neurons of the neural network are pulse neurons or non-pulse neurons. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0038] Figure 1 The overall block diagram of a brain-like visual neural network with forward learning and meta-learning functions provided by the present invention;
[0039] Figure 2 A schematic diagram of an input-side attention control unit and an output-side attention control unit in a composite feature encoding module of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention;
[0040] Figure 3 A schematic diagram of a position encoding unit in a composite feature encoding module of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention;
[0041] Figure 4 It is a topological schematic diagram of an input-side attention control unit, a concrete feature encoding unit, and an abstract feature encoding unit of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention;
[0042] Figure 5 A topological schematic diagram of an input-side attention control unit, a concrete feature encoding unit, an abstract feature encoding unit, and an output-side attention control unit of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention;
[0043] Figure 6 A schematic diagram of the topology of position coding neurons in the corresponding subspace of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention;
[0044] Figure 7 A schematic diagram of the topology of position coding neurons in the corresponding area of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention;
[0045] Figure 8 A schematic diagram of the receptive field projection relationship of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention;
[0046] Fig. 9 A schematic diagram of a forward neural pathway and a reverse neural pathway of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention;
[0047] Fig.10 Schematic diagram of the central-peripheral topological structure of a brain-like visual neural network with forward learning and meta-learning functions in an embodiment of the present invention. DETAILED DESCRIPTION
[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0049] See attached Figure 1 , the embodiment of the present invention discloses a brain-like visual neural network with forward learning and meta-learning functions, including: a plurality of (e.g., 1 to 2) primary feature encoding modules 1, a plurality of (e.g., 3 to 3000) composite feature encoding modules 2;
[0050] Each module includes multiple neurons;
[0051] The neurons include primary feature encoding neurons 10 , concrete feature encoding neurons 210 , and abstract feature encoding neurons 220 .
[0052] The primary feature encoding module 1 includes a plurality of (eg, 2 million) primary feature encoding neurons 10 for encoding primary visual feature information.
[0053] The composite feature encoding module 2 includes a concrete feature encoding unit 21 and an abstract feature encoding unit 22 .
[0054] The concrete feature encoding unit 21 includes a plurality of (eg, 100,000) concrete feature encoding neurons 210 for encoding concrete visual feature information.
[0055] The abstract feature encoding unit 22 includes a plurality of (eg, 100,000) abstract feature encoding neurons 220 for encoding abstract visual feature information.
[0056] In the expression, if a unidirectional connection is formed between neuron A and neuron B, it means a unidirectional connection of A->B; if a bidirectional connection is formed between neuron A and neuron B, it means a bidirectional connection of A<->B (or A->B and A<-B);
[0057] If there is a unidirectional connection between neuron A and neuron B, A is called the direct upstream neuron of neuron B, and neuron B is called the direct downstream neuron of neuron A; if there is a bidirectional connection between neuron A and neuron B, A and neuron B are called the direct upstream neuron and the direct downstream neuron of each other;
[0058] If there is no connection between neuron A and neuron B, but there is a connection pathway between the two through several other neurons, such as A->C->…->D->B, then neuron A is called the indirect upstream neuron of neuron B, neuron B is called the indirect downstream neuron of neuron A, and neuron D is called the direct upstream neuron of neuron B.
[0059] The excitatory connection is: when the upstream neuron of the excitatory connection emits, a non-negative input is provided to the downstream neuron through the excitatory connection.
[0060] The inhibitory connection is: when the upstream neuron of the inhibitory connection is fired, a non-positive input is provided to the downstream neuron through the inhibitory connection.
[0061] A number (eg, 10,000) of the primary feature encoding neurons 10 respectively form unidirectional or bidirectional excitatory / inhibitory connections with other number (eg, 1-20) of the primary feature encoding neurons 10 .
[0062] A plurality (e.g., 500,000 to 1,000,000) of the primary feature coding neurons 10 respectively form unidirectional or bidirectional excitatory / inhibitory connections with a plurality (e.g., 10 to 1,000) of the concrete feature coding neurons 210 or a plurality (e.g., 10-1,000) of the abstract feature coding neurons 220 located in at least one (e.g., 2) of the composite feature coding modules 2.
[0063] Several (e.g., 50,000) concrete feature coding neurons 210 located in the same composite feature coding module 2 respectively form unidirectional or bidirectional excitatory / inhibitory connections with several (e.g., 5,000) abstract feature coding neurons 220 located in the same composite feature coding module 2.
[0064] Several (e.g., 50,000) of the concrete feature coding neurons 210 and the abstract feature coding neurons 220 in several (e.g., 3 to 3,000) of the composite feature coding modules 2 respectively form unidirectional or bidirectional excitatory / inhibitory connections with several (e.g., 2,000) of the concrete feature coding neurons 210 and the abstract feature coding neurons 220 in other several (e.g., 1 to 300) of the composite feature coding modules 2.
[0065] The neural network caches and encodes information through the firing of the neurons, and encodes, stores, and transmits information through the connections between the neurons.
[0066] Input a picture or video stream, multiply the R, G, B pixel values of each pixel of each frame of the picture by weights and input them into a number (such as 2-30) of the primary feature coding neurons 10, so that the primary feature coding neurons 10 are activated.
[0067] For several of the neurons, their membrane potentials are calculated to determine whether to fire. If so, the membrane potentials of the downstream neurons are accumulated to determine whether to fire, so that the firing is propagated in the neural network; the weight of the connection between the upstream neurons and the downstream neurons is a constant, or is dynamically adjusted through the synaptic plasticity process.
[0068] The working process of the neural network includes: forward memory process, memory triggering process, information aggregation process, directional information aggregation process, information transcription process, memory forgetting process, memory self-consolidation process, information component adjustment process, reinforcement learning process, novelty signal modulation process, and supervised learning process.
[0069] The synaptic plasticity process includes a monopolar upstream release-dependent synaptic plasticity process, a monopolar downstream release-dependent synaptic plasticity process, a monopolar upstream and downstream release-dependent synaptic plasticity process, a monopolar upstream pulse-dependent synaptic plasticity process, a monopolar downstream pulse-dependent synaptic plasticity process, a monopolar pulse time-dependent synaptic plasticity process, an asymmetric bipolar pulse time-dependent synaptic plasticity process, and a symmetric bipolar pulse time-dependent synaptic plasticity process.
[0070] Mapping a number of the neurons to corresponding labels as outputs, for example, mapping 100,000 abstract feature encoding neurons 220 of the high-level information channel to corresponding labels as outputs.
[0071] In this embodiment, several neurons of the neural network are pulse neurons or non-pulse neurons.
[0072] In this embodiment, all primary feature encoding neurons 10, concrete feature encoding neurons 210, abstract feature encoding neurons 220, and intermediate neurons are pulse neurons.
[0073] For example, one way to implement a spiking neuron is to use a leaky integral spiking neuron (LIF neuron model); one way to implement a non-spiking neuron is to use an artificial neuron in a deep neural network (for example, using a ReLU activation function).
[0074] In this embodiment, several neurons of the neural network are self-exciting neurons; the self-exciting neurons include conditional self-exciting neurons and unconditional self-exciting neurons;
[0075] If the conditionally self-exciting neuron is not excited by external input in the first preset time interval, it will self-excite according to probability P;
[0076] The membrane potential of the unconditionally self-exciting neuron is automatically accumulated gradually in the absence of external input. When the membrane potential reaches a threshold, the unconditionally self-exciting neuron is excited and the membrane potential is restored to the resting potential to restart the accumulation process.
[0077] In this embodiment, an implementation method of an unconditional self-excitation neuron is:
[0078] Step m1: membrane potential Vm=Vm+Vc;
[0079] Step m2: weighted sum of all inputs and add them to Vm;
[0080] Step m3: If Vm>=threshold, the unconditional self-excitation neuron is excited, and Vm=Vrest;
[0081] Repeat steps m1 to m3;
[0082] Vm is the membrane potential, Vc is the cumulative constant, Vrest is the resting potential, and threshold is the threshold value;
[0083] For example, let Vc=5mV, Vrest=-70mV, threshold=-25mV.
[0084] In this embodiment, if the conditionally self-exciting neuron is not excited by external input within a first preset time interval (for example, configured as 10 minutes), it will self-excite according to probability P;
[0085] The conditionally self-exciting neuron records any one or more of the following information:
[0086] 1) The time interval from the last excitation,
[0087] 2) Recent average issuance rate,
[0088] 3) Duration of the most recent stimulus,
[0089] 4) Total number of excitations,
[0090] 5) The total number of times the synaptic plasticity process of each input connection has been executed recently,
[0091] 6) The total number of times the synaptic plasticity process of each output connection has been executed recently,
[0092] 7) The total weight change of each recent input connection,
[0093] 8) The total weight change of each outgoing connection recently.
[0094] In this embodiment, the calculation rule of the probability P includes any one or more of the following:
[0095] 1) P is positively correlated with the time interval from the last excitation.
[0096] 2) P is positively correlated with the recent average issuance rate.
[0097] 3) P is positively correlated with the duration of the most recent stimulation.
[0098] 4) P is positively correlated with the total number of excitations.
[0099] 5) P is positively correlated with the total number of synaptic plasticity processes executed by the most recent input connections.
[0100] 6) P is positively correlated with the total number of synaptic plasticity processes executed recently for each output connection.
[0101] 7) P is positively correlated with the total weight change of the most recent input connections.
[0102] 8) P is positively correlated with the total weight change of the most recent output connections.
[0103] 9) P is positively correlated with the weighted average of all input connections.
[0104] 10) P is positively correlated with the total modulus of the weights of all input connections.
[0105] 11) P is positively correlated with the total number of all input connections.
[0106] 12) P is positively correlated with the total number of all output connections.
[0107] In this embodiment, let P = min(1, a*Tinterval^2+b*Fr+c*Nin_plasticity+Bias); where: a, b, c are coefficients, Tinterval is the time interval from the last excitation, Fr is the recent average firing rate, Nin_plasticity is the total number of executions of the synaptic plasticity process of each recent input connection, and Bias is the bias.
[0108] In this embodiment, the calculation rules of the activation intensity or firing rate Fs of the conditionally self-exciting neuron during self-excitation include any one or more of the following:
[0109] 1) Fs = Fsd, Fsd is the default excitation frequency,
[0110] 2) Fs is negatively correlated with the time interval from the last excitation.
[0111] 3) Fs is positively correlated with the recent average firing rate.
[0112] 4) Fs is positively correlated with the duration of the most recent excitation.
[0113] 5) Fs is positively correlated with the total number of excitations.
[0114] 6) Fs is positively correlated with the total number of synaptic plasticity processes executed by the most recent input connections.
[0115] 7) Fs is positively correlated with the total number of synaptic plasticity processes executed by the most recent output connections.
[0116] 8) Fs is positively correlated with the total weight change of the most recent input connections.
[0117] 9) Fs is positively correlated with the total weight change of the most recent output connections.
[0118] 10) Fs is positively correlated with the weighted average of all input connections.
[0119] 11) Fs is positively correlated with the total modulus of the weights of all input connections.
[0120] 12) Fs is positively correlated with the total number of all input connections.
[0121] 13) Fs is positively correlated with the total number of all output connections;
[0122] If the conditionally self-exciting neuron is a pulse neuron, P is the probability of currently emitting a series of pulses, if it is emitted, the emission rate is Fs, if it is not emitted, the emission rate is 0;
[0123] If the conditionally self-exciting neuron is a non-pulse neuron, P is the probability of current excitation, if activated, the activation intensity is Fs, if not activated, the activation intensity is 0.
[0124] In this embodiment, 500,000 to 1,000,000 of the primary feature encoding neurons 10, 10,000,000 of the concrete feature encoding neurons 210, 10,000,000 of the abstract feature encoding neurons 220, and 1,000,000 of the input-side attention regulation neurons 230 use the conditional self-exciting neurons; 500,000 to 1,000,000 of the primary feature encoding neurons 10 use unconditional self-exciting neurons.
[0125] In this embodiment, each neuron and each connection (including neuron-neuron connection and synapse-synapse connection) can be represented by a vector or a matrix, and the operation of the neural network is expressed as a vector or a matrix operation; for example, the same type of parameters in each neuron and each connection (such as the firing rate of the neuron and the weight of the connection) are flattened into a vector or a matrix, and the signal propagation of the neural network can be expressed as a dot product operation of the neuron firing rate vector and the connection weight vector (i.e., the weighted sum of the inputs).
[0126] In another embodiment, each neuron and each connection (including neuron-neuron connection and synapse-synapse connection) can also be implemented in an object-oriented manner; for example, they are respectively implemented as an object (object in object-oriented programming), and the operation of the neural network is manifested as the call of the object and the information transfer between objects.
[0127] In another embodiment, the neural network may also be implemented in firmware (such as FPGA) or ASIC (such as a neuromorphic chip).
[0128] In another embodiment, several connections of the neural network may be replaced by convolution operations; for example, all connections between each primary feature encoding neuron 10 and each image feature encoding neuron 210 may be replaced by convolution operations, and a signal projection relationship having one or more receptive fields may also be generated. The projection relationship of the receptive field can be found in the attached Figure 8.
[0129] See attached Figure 2 , 3 , 4, 5, in a further improved embodiment, the composite feature encoding module 2 may further include an input side attention control unit and an output side attention control unit;
[0130] The neurons also include an input-side attention regulation neuron 230 and an output-side attention regulation neuron 240;
[0131] The input-side attention control unit includes a plurality of (e.g., 100,000) input-side attention control neurons 230;
[0132] The output-side attention regulation unit includes a plurality of (eg, 100,000) output-side attention regulation neurons 240 .
[0133] A number (e.g., 50,000) of the input-side attention regulation neurons 230 can respectively receive unidirectional or bidirectional excitatory / inhibitory connections from a number (e.g., 100,000 to 10,000) of the primary feature encoding neurons 10.
[0134] Each of the input-side attention regulation neurons 230 forms a unidirectional or bidirectional excitatory connection with a number (e.g., 1 to 1,000) of the concrete feature encoding neurons 210 / abstract feature encoding neurons 220 of the composite feature encoding module 2 in which it is located.
[0135] Each of the input-side attention regulation neurons 230 receives unidirectional or bidirectional excitatory connections from a number (e.g., 100,000 to 10,000) of the concrete feature coding neurons 210 / abstract feature coding neurons 220 / output-side attention regulation neurons 240 from other composite feature coding modules 2.
[0136] Each of the input-side attention regulation neurons 230 can also form a unidirectional or bidirectional excitatory connection with several other (eg, 1,000) input-side attention regulation neurons 230 .
[0137] Each of the output-side attention regulation neurons 240 forms a unidirectional or bidirectional excitatory connection with a number (e.g., 1,000 to 10,000) of the concrete feature coding neurons 210 / abstract feature coding neurons 220 / input-side attention regulation neurons 230 located in other composite feature coding modules 2.
[0138] Each of the output-side attention regulation neurons 240 receives unidirectional or bidirectional excitatory connections from a number (e.g., 1 to 1,000) of the concrete feature encoding neurons 210 / abstract feature encoding neurons 220 of the composite feature encoding module 2 in which it is located.
[0139] Each of the output-side attention regulation neurons 240 can also form a unidirectional or bidirectional excitatory connection with several other (eg, 1,000) output-side attention regulation neurons 240 .
[0140] Each of the input-side attention regulation neurons 230 may have an input-side attention control terminal 31 ; each of the output-side attention regulation neurons 240 may have an output-side attention control terminal 32 .
[0141] The working process of the neural network also includes an active attention process and an automatic attention process.
[0142] The active attention process is: adjusting the activation intensity or firing rate or pulse firing phase of each input-side attention regulation neuron 230 by the strength of the attention control signal applied at the input-side attention control terminal 31 (the amplitude can be positive, negative, or 0), thereby controlling the information entering the corresponding concrete feature coding unit 21 and the abstract feature coding unit 22, and adjusting the size and proportion of each information component; or, adjusting the activation intensity or firing rate or pulse firing phase of each output-side attention regulation neuron 240 by the strength of the attention control signal applied at the output-side attention control terminal 32 (the amplitude can be positive, negative, or 0), thereby controlling the information output from the corresponding concrete feature coding unit 21 and the abstract feature coding unit 22, and adjusting the size and proportion of each information component.
[0143] The automatic attention process is: when several neurons connected to the input side attention regulation neurons 230 are activated, these input side attention regulation neurons 230 are easier to activate, so that the relevant information components are easier to input into the corresponding concrete feature encoding unit 21 and the abstract feature encoding unit 22; or, when several neurons connected to the output side attention regulation neurons 240 are activated, these output side attention regulation neurons 240 are easier to activate, so that the relevant information components are easier to output from the corresponding concrete feature encoding unit 21 and the abstract feature encoding unit 22.
[0144] In a further improved embodiment, the neural network includes one or more information channels;
[0145] The working process of the neural network also includes the process of automatic formation of information channels.
[0146] The automatic formation process of the information channel is: by performing any one or several of the forward memory process, memory triggering process, information aggregation process, directional information aggregation process, information transcription process, memory forgetting process, memory self-consolidation process, information component adjustment process, active attention process, and automatic attention process, the connection relationship and weight between each of the neurons are adjusted so that the neural network forms one or more information channels, each of which encodes one or more information components; there may be overlaps between the information channels.
[0147] The neural network can also form one or more information channels by presetting initial connection relationships and initial parameters (such as connection weights, thresholds of the neurons, initial membrane potentials of the neurons, and initial time constants of the neurons), and each information channel encodes one or more preset information components.
[0148] In this embodiment, the information channel includes a primary information channel;
[0149] The primary information channel is: all the primary feature encoding neurons 10 and their connections constitute the primary information channel.
[0150] The primary information channels include a primary contrast information channel, a primary orientation information channel, a primary edge information channel, and a primary color block information channel.
[0151] See attached Fig.10 The primary contrast information channel is: select a number of adjacent pixels in the input image as central area pixels, select a number of pixels around the central area pixels as peripheral area pixels, multiply the pixel values of each of the central area pixels and the peripheral area pixels by weights and input them into a number of the primary feature encoding neurons 10, that is, forming a central-peripheral topological structure. These primary feature encoding neurons 10 and their connections constitute one or more primary contrast information channels.
[0152] In the primary information channel, one or more numbers of pixels, positions covering the screen space, and adjacent pixels of area are selected, and one or more pixel values of these pixels are multiplied by one or more weights respectively. It is also possible to form several primary orientation information channels, primary edge information channels, primary color block information channels, or a combination thereof with one or more receptive fields.
[0153] For example, the R, G, and B pixel values of one of the central area pixels are multiplied by negative weights (such as -2), negative weights (such as -2), and positive weights (such as +4), respectively, and input into several (such as 1-2) primary feature coding neurons 10; the R, G, and B pixel values of the surrounding area pixels at the upper left (such as 1) and the lower right (such as 1) are multiplied by negative weights (such as -2), negative weights (such as -2), and positive weights (such as +4), respectively, and the R, G, and B pixel values of the surrounding area pixels at the top (such as 1), bottom (such as 1), left (such as 1), right (such as 1), upper right (such as 1), and lower left (such as 1) are multiplied by positive weights (such as +2), positive weights (such as +2), and negative weights (such as -4), respectively, and input into these primary feature coding neurons 10. These primary feature coding neurons 10 and their connections constitute an information channel with a blue-yellow contrast and a 3x3 receptive field that is sensitive to a 45° orientation from the upper left to the lower right.
[0154] For example, in a 15x15 pixel area, the R, G, and B values of each pixel are multiplied by negative weight, negative weight, and positive weight, respectively, and input into a number (e.g., 1-2) of the primary feature coding neurons 10. These primary feature coding neurons 10 and their connections constitute a color block information channel that is sensitive to blue.
[0155] The primary contrast information channel includes a light-dark contrast information channel, a dark-light contrast information channel, a red-green contrast information channel, a green-red contrast information channel, a yellow-blue contrast information channel, and a blue-yellow contrast information channel.
[0156] The light-dark contrast information channel is: the R, G, and B pixel values of each (e.g., 9) central area pixel are multiplied by positive weights (e.g., +1) and input into a number (e.g., 1 to 10) of primary feature coding neurons 10, and the R, G, and B pixel values of each (e.g., 72) surrounding area pixel are multiplied by negative weights (e.g., -1) and input into these primary feature coding neurons 10. These primary feature coding neurons 10 and their connections constitute the light-dark contrast information channel.
[0157] The dark-light contrast information channel is: the R, G, and B pixel values of each (e.g., 9) central area pixel are multiplied by negative weights (e.g., -1) and input into a number (e.g., 1 to 10) of primary feature coding neurons 10, and the R, G, and B pixel values of each (e.g., 72) surrounding area pixel are multiplied by positive weights (e.g., +1) and input into these primary feature coding neurons 10. These primary feature coding neurons 10 and their connections constitute the dark-light contrast information channel.
[0158] See attached Fig.10, the red-green contrast information channel is: the R, G, B pixel values of each (e.g., 4) central area pixel are multiplied by positive weights (e.g., +2), negative weights (e.g., -2), and positive weights (e.g., +1) and input into a number (e.g., 1 to 10) of primary feature coding neurons 10, and the R, G, B pixel values of each (e.g., 32) surrounding area pixel are multiplied by negative weights (e.g., -2), positive weights (e.g., +2), and positive weights (e.g., +1) and input into these primary feature coding neurons 10, and these primary feature coding neurons 10 and their connections constitute the red-green contrast information channel.
[0159] The green-red contrast information channel is: the R, G, and B pixel values of each (e.g., 4) central area pixel are multiplied by negative weights (e.g., -2), positive weights (e.g., +2), and positive weights (e.g., +1) and input into a number (e.g., 1 to 10) of primary feature coding neurons 10, and the R, G, and B pixel values of each (e.g., 32) surrounding area pixel are multiplied by positive weights (e.g., +2), negative weights (e.g., -2), and positive weights (e.g., +1) and input into these primary feature coding neurons 10. These primary feature coding neurons 10 and their connections constitute the green-red contrast information channel.
[0160] The yellow-blue contrast information channel is: the R, G, and B pixel values of each (e.g., 4) central area pixel are multiplied by positive weights (e.g., +2), positive weights (e.g., +2), and negative weights (e.g., -4) and input into a number (e.g., 1 to 10) of primary feature coding neurons 10, and the R, G, and B pixel values of each (e.g., 32) surrounding area pixel are multiplied by negative weights (e.g., -2), negative weights (e.g., -2), and positive weights (e.g., +4) and input into these primary feature coding neurons 10. These primary feature coding neurons 10 and their connections constitute the yellow-blue contrast information channel.
[0161] The blue-yellow contrast information channel is: the R, G, and B pixel values of each (e.g., 4) central area pixel are multiplied by negative weights (e.g., -2), negative weights (e.g., -2), and positive weights (e.g., +4) and input into a number (e.g., 1 to 10) of primary feature coding neurons 10, and the R, G, and B pixel values of each (e.g., 32) surrounding area pixel are multiplied by positive weights (e.g., +2), positive weights (e.g., +2), and negative weights (e.g., -4) and input into these primary feature coding neurons 10. These primary feature coding neurons 10 and their connections constitute the blue-yellow contrast information channel.
[0162] Specifically, the primary visual feature information includes light-dark contrast information, dark-light contrast information, red-green contrast information, green-red contrast information, yellow-blue contrast information, blue-yellow contrast information, primary edge information, primary orientation information, receptive field information, and color block information.
[0163] In a further improved embodiment, the primary information channel also includes a primary optical flow information channel;
[0164] The primary optical flow information channel is: calculating the optical flow for each pixel in the input image, obtaining the direction value and speed value of the optical flow movement, combining different direction values and speed values, and multiplying them by weights and inputting them into a number (e.g., 1 to 10) of the primary feature encoding neurons 10, and these primary feature encoding neurons 10 and their connections constitute the primary optical flow information channel;
[0165] The primary visual feature information also includes optical flow information.
[0166] See attached Figure 3 , Attachment Figure 6 , Attachment Figure 7 In a further improved embodiment, the composite feature encoding module 2 may further include a plurality of (eg, 1 to 10) position encoding units 25;
[0167] The neurons also include position encoding neurons 250;
[0168] The position encoding unit 25 includes a plurality of (eg, 1,000 to 10,000) position encoding neurons 250 for encoding position information (of visual features relative to the picture space or relative to other visual features).
[0169] Each of the position encoding units 25 corresponds to a plurality of subspaces in the picture space, and the subspaces may have intersections.
[0170] Each of the position coding neurons 250 corresponds to each region corresponding to the position in each subspace corresponding to the position coding unit 25, and receives unidirectional or bidirectional excitatory connections of a number (e.g., 1 to 10,000) of neurons (e.g., primary feature coding neurons 10) whose receptive fields are in these regions; the projection relationship of the receptive fields can be found in the attached Figure 8 .
[0171] Several (eg, each) of the position coding neurons 250 form unidirectional or bidirectional excitatory connections with other position coding neurons 250 corresponding to the same region.
[0172] A number (e.g., 1,000 to 50,000) of the position coding neurons 250 can also form unidirectional or bidirectional excitatory connections with a number (e.g., 1 to 1,000) of the input side attention regulation neurons 230 / output side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 located in the composite feature coding module 2 where they are located.
[0173] A number (e.g., 1,000 to 50,000) of the position coding neurons 250 can also form unidirectional or bidirectional excitatory connections with a number (e.g., 1 to 1,000) of the input-side attention regulation neurons 230 / concrete feature coding neurons 210 / abstract feature coding neurons 220 located in other composite feature coding modules 2.
[0174] For example, in the attached Figure 6 In the figure, the position coding neurons 250A, 250B, 250C, and 250D correspond to some subspaces / regions in the picture space respectively, and the subspaces / regions corresponding to them have intersections with each other; the position coding neurons 250A, 250B, 250C, and 250D correspond to the same regions as the position coding neurons 250E, 250F, 250G, and 250H respectively, and bidirectional excitatory connections are formed between them.
[0175] For example, in the attached Figure 7 In the figure, the position encoding neuron 250Y receives excitatory connections from the primary feature encoding neurons 10E, 10F, etc. in the regions corresponding to the positions in the four subspaces.
[0176] In a further improved embodiment, the information channel also includes an intermediate information channel;
[0177] The intermediate information channel includes an intermediate position information channel;
[0178] The intermediate position information channel is: through the automatic formation process of the information channel, or by presetting the initial connection relationship and initial parameters, the proportion of the total weight of the connection from the position coding neuron 250 and the neurons encoding the position information in the total weight of some or all input connections of all the input side attention regulation neurons 230 / output side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 in a number (such as 1 to 10) of the composite feature coding modules 2 reaches or exceeds a first preset proportion (such as configured as 30%), and the connection weights from the position coding neurons 250 and the neurons encoding the position information are combined in multiple proportions, so that these input side attention regulation neurons 230 / output side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 respectively have one or more receptive fields, respectively encode one or more position information, and together with the position coding neurons 250 constitute the intermediate position information channel.
[0179] Since the intermediate position information channel includes neurons encoding the position information of visual features, an explicit position encoding method is adopted.
[0180] In this embodiment, the intermediate information channel also includes an intermediate visual feature information channel;
[0181] The intermediate visual feature information channel is: through the automatic formation process of the information channel, or by presetting the initial connection relationship and initial parameters, the proportion of the total connection weight of neurons from the primary information channel in the total weight of some or all input connections of the input side attention regulation neurons 230 / output side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 of a number (such as 10 to 2000) of the composite feature coding modules 2 reaches or exceeds a second preset proportion (such as configured as 60%), and the connection weights of each neuron from each area and position in the corresponding picture space of the primary information channel and the intermediate information channel are combined in multiple proportions, so that these input side attention regulation neurons 230 / output side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 respectively have one or more receptive fields, respectively encode one or more intermediate visual feature information, and together constitute the intermediate visual feature information channel.
[0182] Specifically, the intermediate visual feature information includes composite color contrast information, composite light-dark contrast information, composite orientation information, composite edge information, area information, and motion information.
[0183] For example, 1 to 10 of the composite feature coding modules 2 are selected, and 80% of the input-side attention regulation neurons 230 / output-side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 are made to receive unidirectional excitatory connections from neurons of the two primary orientation information channels respectively. Then, the neurons of these composite feature coding modules 2 encode the integration of the primary orientation information (one is the horizontal right direction, and the other is the direction of 10° from the upper left to the lower right, and the receptive fields are both 9x9) encoded by the two primary orientation information channels, that is, the integrated orientation information (the direction interval from the horizontal right to the upper left to the lower right 10°, and the receptive field is 9x9).
[0184] Since the neurons of the intermediate visual feature information channel directly accept connections from the neurons of the primary information channel, these connection relationships correspond to various areas and positions in the picture space to form a receptive field, that is, an implicit position encoding method is adopted.
[0185] In this embodiment, the information channel also includes an advanced information channel;
[0186] The high-level information channel is: through the automatic formation process of the information channel, or through the preset initial connection relationship and initial parameters, the proportion of the total connection weight of neurons from the intermediate information channel in the total weight of some or all input connections of the input side attention regulation neurons 230 / output side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 in a number (such as 10 to 2000) of the composite feature coding modules 2 reaches or exceeds a third preset ratio (such as configured as 40%), and the connection weights of each neuron in each area and position in the corresponding picture space of the primary information channel, the intermediate information channel, and the high-level information channel are combined in multiple proportions, so that these input side attention regulation neurons 230 / output side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 respectively have one or more receptive fields, respectively encode one or more high-level visual feature information, and together constitute the high-level information channel.
[0187] Specifically, the high-level visual feature information includes contour information, texture information, brightness information, transparency information, shape and position information, compound motion information, and objectification information.
[0188] The objectified information is the identified object (which may be an instance or a category), and each object may have a name, such as "apple", "banana", and "car".
[0189] For example, select 1 to 10 of the composite feature coding modules 2, and make 80% of the input side attention regulation neurons 230 / output side attention regulation neurons 240 / concrete feature coding neurons 210 / abstract feature coding neurons 220 receive unidirectional excitatory connections from neurons of multiple intermediate information channels respectively, and these intermediate information channels encode composite edge information and position information, then the neurons of these composite feature coding modules 2 encode shape and position information.
[0190] In this embodiment, a basic working process of the neural network, its modules or units is: select a number of starting neurons, source neurons, and target neurons from a number of candidate neurons (in one or some modules or sub-modules), and make a number of the starting neurons generate a firing distribution and remain activated for a preset time or operation cycle, so that the connections between the neurons involved in the working process adjust the weights through the synaptic plasticity process.
[0191] The firing distribution is: a plurality of the neurons generate the same or different activation intensities, firing rates, and pulse phases;
[0192] For example, neuron A, neuron B, and neuron C produce activation intensities of amplitudes 2, 5, and 9, respectively, or produce firing rates of 0.4 Hz, 50 Hz, and 20 Hz, respectively, or produce pulse phases of 100 ms, 300 ms, and 150 ms, respectively.
[0193] The process of selecting an initiating neuron, a source neuron or a target neuron from a number of candidate neurons includes any one or more of the following: selecting the first Kf1 neuron with the smallest total modulus of weights of some or all input connections, selecting the first Kf2 neuron with the smallest total modulus of weights of some or all output connections, selecting the first Kf3 neuron with the largest total modulus of weights of some or all input connections, selecting the first Kf4 neuron with the largest total modulus of weights of some or all output connections, selecting the first Kf5 neuron with the largest activation intensity or firing rate or the first to start firing, selecting the first Kf6 neuron with the smallest activation intensity or firing rate or the latest to start firing (including no firing), selecting the first Kf7 neuron with the longest time since the last firing, selecting the first Kf8 neuron with the closest time to the last firing, selecting the first Kf9 neuron with the longest time since the last input connection or output connection underwent synaptic plasticity process, and selecting the first Kf10 neuron with the closest time since the last input connection or output connection underwent synaptic plasticity process.
[0194] The method of causing several of the neurons to generate a firing distribution and maintain activation for a preset period (e.g., 200ms to 2s) may be to input samples (pictures or video streams), directly activate several neurons in the neural network, self-excite several neurons in the neural network, or propagate the existing activation states of several neurons in the neural network in the neural network to activate several neurons (e.g., the starting neurons).
[0195] See attached Fig. 9 , in this embodiment, the neural network includes a forward neural pathway and a reverse neural pathway;
[0196] The forward neural pathway and the reverse neural pathway are respectively: a plurality of the primary feature coding modules 1 / composite feature coding modules 2 are cascaded in a first preset order, a neural pathway formed by a plurality of the neurons cascaded in the first preset order is used as the forward neural pathway, and a neural pathway formed by a plurality of the neurons cascaded in the opposite direction of the first preset order is used as the reverse neural pathway.
[0197] Specifically, the first preset order is primary information channel, intermediate information channel, and advanced information channel; the forward neural pathway is a neural pathway composed of neurons of several primary information channels, intermediate information channels, and advanced information channels through a bottom-up cascade (i.e., following the first preset order), which is mainly involved in recognizing external input information (pictures or video streams) and the forward learning process; the reverse neural pathway is a neural pathway composed of neurons of several advanced information channels, intermediate information channels, and primary information channels through a top-down cascade (i.e., against the first preset order), which is mainly involved in the pattern completion process, directional initiation process, association process or imagination process.
[0198] In each primary feature encoding module 1 / composite feature encoding module 2, several neurons constituting the forward neural pathway can form unidirectional or bidirectional excitatory / inhibitory connections with several neurons constituting the reverse neural pathway.
[0199] In this embodiment, the working process of the neural network also includes a directional start process;
[0200] The directional starting process includes a forward starting process and a reverse starting process.
[0201] The forward start process is:
[0202] Step o1: Select several neurons in the forward neural pathway as starter neurons;
[0203] Step o2: enabling each of the oscillating neurons to generate a firing distribution and maintain activation for a third preset period Tfprime;
[0204] Step o3: a number of the neurons in the reverse neural pathway that receive excitatory connections from the initiating neuron receive non-negative inputs to be more easily activated;
[0205] Step o4: Some of the neurons in the reverse neural pathway that receive inhibitory connections from the excitatory neurons receive non-positive inputs to make them more difficult to activate.
[0206] The reverse startup process is:
[0207] Step n1: Select several neurons in the reverse neural pathway as starter neurons;
[0208] Step n2: causing each of the oscillating neurons to generate a firing distribution and remain activated for a tenth preset period Tbprime;
[0209] Step n3: a number of the neurons in the forward neural pathway that receive excitatory connections from the energizing neuron receive non-negative inputs to be more easily activated;
[0210] Step n4: Some of the neurons in the forward neural pathway that receive inhibitory connections from the excitatory neuron receive non-positive inputs to make them more difficult to activate.
[0211] For example, Tfprime and Tbprime are configured to 5 seconds.
[0212] The directional activation process can be used for visual search. When performing visual search, several of the neurons in the reverse neural pathway that encode the searched information components are used as starting neurons and generate a discharge distribution that represents the searched information components and remain activated for the tenth preset period Tbprime (such as configured to be 5 seconds). Several of the neurons in the forward neural pathway that encode the searched information components are easier to activate, while several of the neurons that do not encode the searched information components are inhibited. When the searched information components appear in the external input information (picture or video stream), they are easier to be identified, and irrelevant information components are filtered out.
[0213] See attached Fig. 9 For example, the neural network can be configured to be composed of a primary feature coding module 1A, a composite feature coding module 2A, a composite feature coding module 2B, and a composite feature coding module 2C; these four modules respectively include primary feature coding neurons 10A, 10B, and 10C, concrete feature coding neurons 210A, 210B, and 210C, concrete feature coding neurons 210D, 210E, and 210F, and concrete feature coding neurons 210G, 210H, and 210I.
[0214] The first preset order can be configured as the order of primary feature coding module 1A, composite feature coding module 2A, composite feature coding module 2B, and composite feature coding module 2C; primary feature coding neurons 10A, concrete feature coding neurons 210A, concrete feature coding neurons 210D, and concrete feature coding neurons 210G follow the first preset order through a unidirectional excitatory connection cascade to form a forward neural pathway A; primary feature coding neurons 10B, concrete feature coding neurons 210B, concrete feature coding neurons 210E, and concrete feature coding neurons 210H follow the first preset order through a unidirectional excitatory connection cascade to form a forward neural pathway B; primary feature coding neurons 10C, concrete feature coding neurons 210C, concrete feature coding neurons 210F, and concrete feature coding neurons 210I follow the first preset order through a unidirectional excitatory connection cascade to form a reverse neural pathway C.
[0215] The primary feature encoding neuron 10C forms a bidirectional excitatory connection and a bidirectional inhibitory connection with the primary feature encoding neurons 10A and 10B respectively.
[0216] The concrete feature encoding neuron 210C forms a bidirectional excitatory connection and a bidirectional inhibitory connection with the concrete feature encoding neurons 210A and 210B respectively.
[0217] The concrete feature encoding neuron 210F forms a bidirectional excitatory connection and a bidirectional inhibitory connection with the concrete feature encoding neurons 210D and 210E respectively.
[0218] The concrete feature encoding neuron 210I forms a bidirectional excitatory connection and a bidirectional inhibitory connection with the concrete feature encoding neurons 210G and 210H respectively.
[0219] Thus, the reverse neural pathway C promotes the forward neural pathway A and inhibits the forward neural pathway B.
[0220] In this embodiment, the neurons also include interneurons;
[0221] The primary feature coding module 1 and the composite feature coding module 2 respectively include a number (e.g., 1,000 to 10,000) of the intermediate neurons, and the several intermediate neurons respectively form unidirectional inhibitory connections with a number (e.g., 1 to 10,000) of corresponding neurons in the corresponding modules, and the corresponding several neurons in each module respectively form unidirectional excitatory connections with a number (e.g., 1 to 10,000) of corresponding intermediate neurons.
[0222] In this embodiment, the corresponding two or more groups of neurons in each module form inter-group competition (lateral inhibition) through the intermediate neuron. When input is applied, the competing groups of neurons produce different overall activation strengths (or firing rates). The lateral inhibition of the intermediate neuron makes the overall activation strength (or firing rate) stronger and weaker, or the neurons (groups) that start firing first inhibit the neurons (groups) that fire later to form a time difference, thereby ensuring that the information encoding of each group of neurons is independent, decoupled from each other, and automatically grouped, so that the input information in the memory triggering process can trigger the memory information with the highest correlation, and also enable the neurons participating in the directional information aggregation process to be automatically grouped into Ga1, Ga2, Ga3, and Ga4 according to the response (activation strength or firing rate, or the order of firing time).
[0223] In this embodiment, the neuron further includes a differential information decoupling neuron, and the working process of the neural network further includes a differential information decoupling process;
[0224] The differential information decoupling process is:
[0225] Select a number of the input-side attention regulation neurons 230 / output-side attention regulation neurons 240 / concrete feature encoding neurons 210 / abstract feature encoding neurons 220 as target neurons;
[0226] Selecting a number of neurons having unidirectional / bidirectional excitatory connections with the target neuron as concrete information source neurons;
[0227] Selecting several other neurons having unidirectional / bidirectional excitatory connections with the target neuron as abstract information source neurons;
[0228] Each of the concrete information source neurons may have a number of the differential information decoupling neurons matched with it; each of the concrete information source neurons forms a unidirectional excitatory connection with each of the matched differential information decoupling neurons; the differential information decoupling neurons form a unidirectional inhibitory connection with the target neurons or form a unidirectional inhibitory synapse-synapse connection with the connection of the information source neuron input to the target neuron, so that the signal input from the concrete information source neuron to the target neuron is subject to inhibitory regulation by the matched differential information decoupling neuron; the abstract information source neuron forms a unidirectional excitatory connection with the differential information decoupling neuron;
[0229] Each of the differential information decoupling neurons may have a decoupling control signal input terminal; the degree of information decoupling is adjusted by adjusting the signal size (which may be positive, negative, or 0) applied to the decoupling control signal input terminal.
[0230] The weight of the unidirectional excitatory connection between the concrete information source neuron / abstract information source neuron and the matching differential information decoupling neuron is a constant value, or is dynamically adjusted through the synaptic plasticity process.
[0231] In this embodiment, one scheme of the synapse-synapse connection is that connection Sconn1 receives input from several other connections (denoted as Sconn2). When the upstream neuron of connection Sconn1 emits, the value transmitted by connection Sconn1 to the downstream neuron is the weight of connection Sconn1 plus the input value of each connection Sconn2.
[0232] In this embodiment, a certain composite feature encoding module 2 is selected to perform the forward learning process, one group of the input-side attention regulation neurons 230 is selected as the starting neurons, and one group of the concrete feature encoding neurons 210 is selected as the target neurons; when a novel sample (picture or video stream) is input, a plurality of the target neurons are activated and the visual feature information in the sample is encoded into concrete feature information (a type of memory information, i.e., the original feature information component of each object) and stored through the forward learning process.
[0233] Then, the composite feature coding module 2 is ordered to perform the directional information aggregation process, select the same group of input-side attention control neurons 230 as the starter neurons, select the same group of concrete feature coding neurons 210 as the source neurons, and select one group of the abstract feature coding neurons 220 as the target neurons; after completing one or more directional information aggregation processes, the concrete feature information is aggregated into abstract feature information (a type of memory information, that is, the common feature information component between objects), and is encoded and stored through multiple target neurons.
[0234] Then, the composite feature coding module 2 is ordered to perform the differential information decoupling process, select the same group of concrete feature coding neurons 210 as the concrete information source neurons, select the same group of abstract feature coding neurons 220 as the abstract information source neurons, and select multiple output-side attention regulation neurons 240 of the composite feature coding module 2 as target neurons; when the same sample is input again, multiple concrete information source neurons are activated, triggering the concrete feature information encoded by them, and multiple abstract information source neurons are also activated at the same time, triggering the abstract feature information encoded by them; these abstract information source neurons activate the differential information decoupling neurons, thereby inhibiting the signals input to each target neuron by these concrete information source neurons, so that the abstract feature information replaces the original concrete feature information and is input to each target neuron, that is, the information finally output to other composite feature coding modules 2 through the composite feature coding module 2 is abstract feature information.
[0235] Then, the composite feature coding module 2 can be made to perform the information component adjustment process, select the same group of input-side attention regulation neurons 230 as the starter neurons, select the same group of concrete feature coding neurons 210 as the target neurons, and let Kb1 take a smaller value (for example, 1). After completing one or more of the information component adjustment processes, the feature information of each target neuron becomes differential feature information (a type of memory information, that is, an information component that characterizes the differences between objects); at this time, the signals output by these concrete feature coding neurons 210 to the output-side attention regulation neurons 240 are no longer inhibited by the differential information decoupling neurons and can be transmitted to the downstream neural network.
[0236] The whole process can be executed one or more times, so that the concrete feature information is gradually abstracted into abstract feature information, and the difference feature information is retained, forming a sparser code, saving coding and signal transmission bandwidth, and making the generalization ability of the representation of the neural network better (because abstract feature information is formed), and details will not be lost in the process of forming a higher-level representation (because the difference feature information is retained).
[0237] In this embodiment, the forward learning process is:
[0238] Step a1: selecting a number of the neurons as starter neurons;
[0239] Step a2: selecting a number of the neurons as target neurons;
[0240] Step a3: adjusting the weights of the unidirectional excitatory connections between each activated initiator neuron and a plurality of target neurons through the synaptic plasticity process;
[0241] Step a4: Each activated target neuron can establish unidirectional or bidirectional excitatory connections with a number of other target neurons, and can also establish self-circulating excitatory connections with itself, and the weights of these connections are adjusted through the synaptic plasticity process;
[0242] When the weights of the input / output connections of each of the target neurons are adjusted through the synaptic plasticity process, the weights of some or all of the input or output connections may be normalized or not.
[0243] In this embodiment, during the forward learning process, 10,000 of the input-side attention regulation neurons 230 are selected as starter neurons, and 10,000 of the concrete feature encoding neurons 210 / abstract feature encoding neurons 220 are selected as target neurons.
[0244] The forward learning process can quickly encode the visual feature information of each object in the current input image / video stream and store it in the primary feature encoding module 1 / concrete feature encoding unit 21 / abstract feature encoding unit 22, so as to facilitate rapid identification when the same or similar object is seen again, and provide processing materials for the information aggregation process / directional information aggregation process / information component adjustment process, find the clustering centers (i.e. common features) and differentiated features of multiple similar objects, which is the basis of meta-learning.
[0245] In this embodiment, the memory triggering process is: input information (picture or video stream), or directly activate several neurons in the neural network, or self-excite several neurons in the neural network, or propagate the existing activation state of several neurons in the neural network in the neural network. If it causes several neurons in the target area to fire within a second preset period (such as 1s), then the representation of each fired neuron in the target area and its activation intensity or firing rate can be used as the result of the memory triggering process;
[0246] The target area may be any sub-network in the neural network (eg, all abstract feature encoding neurons 220 of a certain composite feature encoding module 2).
[0247] In this embodiment, the memory triggering process can be reflected as a recognition process of input information (pictures or video streams), and each emitted neuron in the target area can be mapped to a number of labels through a number of readout layer neurons as the recognition result; each neuron in the target area forms a unidirectional excitatory or inhibitory connection with a number of the readout layer neurons; each readout layer neuron corresponds to a label, and the higher its activation intensity or emission rate, or the earlier it starts to emit, the higher the correlation between the input information and its corresponding label, and vice versa; for example, each of the labels can be "apple", "car", "grassland", etc.
[0248] In this embodiment, the information aggregation process is:
[0249] Step g1: Selecting a number of the neurons as starter neurons;
[0250] Step g2: selecting a number of the neurons as source neurons;
[0251] Step g3: selecting a number of the neurons as target neurons;
[0252] Step g4: causing each of the oscillating neurons to generate a firing distribution and remain activated for an eighth preset period Tk;
[0253] Step g5: in the eighth preset period Tk, the weights of the unidirectional or bidirectional excitatory / inhibitory connections between each activated starting neuron and a plurality of target neurons are adjusted through the synaptic plasticity process;
[0254] Step g6: In the eighth preset period Tk, the weights of the unidirectional or bidirectional excitatory / inhibitory connections between each activated source neuron and a plurality of target neurons are adjusted through the synaptic plasticity process;
[0255] Step g7: Each execution of steps g1 to g6 is recorded as one iteration, and one or more iterations are performed;
[0256] A number of the target neurons are mapped to corresponding labels as a result of the information aggregation process.
[0257] For example, in the information aggregation process, the eighth preset period Tk is configured to be 100ms to 2 seconds, and 10,000 of the input-side attention regulation neurons 230 of any one of the composite feature coding modules 2 are selected as starter neurons, and 10,000 of the concrete feature coding neurons 210 of the composite feature coding module 2 are selected as source neurons, and 10,000 of the abstract feature coding neurons 220 of the composite feature coding module 2 are selected as target neurons.
[0258] In this embodiment, the directional information aggregation process is:
[0259] Step h1: Selecting a number of the neurons as starter neurons;
[0260] Step h2: selecting a number of the neurons as source neurons;
[0261] Step h3: selecting a number of the neurons as target neurons;
[0262] Step h4: causing each of the oscillating neurons to generate a firing distribution and remain activated for a ninth preset period Ta;
[0263] Step h5: in the ninth preset period Ta, Ma1 of the source neurons are activated, and Ma2 of the target neurons are activated;
[0264] Step h6: In the ninth preset period Ta, the first Ka1 source neurons with the highest activation intensity or the largest firing rate or the first to start firing are recorded as Ga1, and the remaining Ma1-Ka1 activated source neurons are recorded as Ga2;
[0265] Step h7: In the ninth preset period Ta, the first Ka2 target neurons with the highest activation intensity or the largest firing rate or the first firing are recorded as Ga3, and the remaining Ma2-Ka2 activated target neurons are recorded as Ga4;
[0266] Step h8: In the ninth preset cycle Ta, the unidirectional or bidirectional excitatory / inhibitory connections of each source neuron in Ga1 and a plurality of target neurons in Ga3 are respectively subjected to one or more synaptic weight enhancement processes;
[0267] Step h9: In the ninth preset cycle Ta, the unidirectional or bidirectional excitatory / inhibitory connections of each source neuron in Ga1 and a plurality of target neurons in Ga4 are respectively subjected to one or more synaptic weight weakening processes;
[0268] Step h10: In the ninth preset period Ta, the unidirectional or bidirectional excitatory / inhibitory connections between each source neuron in Ga2 and a plurality of target neurons in Ga3 may or may not undergo one or more synaptic weight weakening processes;
[0269] Step h11: In the ninth preset cycle Ta, the unidirectional or bidirectional excitatory / inhibitory connections between each source neuron in Ga2 and a plurality of target neurons in Ga4 may or may not undergo one or more synaptic weight enhancement processes;
[0270] Step h12: In the ninth preset period Ta, each activated starting neuron performs one or more synaptic weight enhancement processes with the unidirectional or bidirectional excitatory / inhibitory connections of several target neurons in Ga3;
[0271] Step h13: In the ninth preset period Ta, each activated starting neuron performs one or more synaptic weight weakening processes with the unidirectional or bidirectional excitatory / inhibitory connections of several target neurons in Ga4;
[0272] Step h14: Each execution of the process from step h1 to step h13 is recorded as one iteration, and one or more iterations are performed;
[0273] In the process of steps h8 to h13, after performing one or more synaptic weight enhancement processes or synaptic weight reduction processes, the weights of some or all input or output connections of each source neuron or target neuron may be normalized or not normalized;
[0274] The synaptic weight enhancement process may adopt the unipolar upstream and downstream release-dependent synaptic enhancement process, or the unipolar pulse time-dependent synaptic enhancement process;
[0275] The synaptic weight weakening process may adopt the unipolar upstream and downstream release-dependent synaptic weakening process, or the unipolar pulse time-dependent synaptic weakening process;
[0276] The synaptic weight enhancement process and the synaptic weight weakening process may also respectively adopt the asymmetric bipolar pulse time-dependent synaptic plasticity process or the symmetric bipolar pulse time-dependent synaptic plasticity process;
[0277] The representation of each of the target neurons may be mapped to a corresponding label as an output as a result of the directed information aggregation process of the representation of each of the source neurons;
[0278] Ma1 and Ma2 are positive integers, Ka1 is a positive integer not exceeding Ma1, and Ka2 is a positive integer not exceeding Ma2.
[0279] For example, in the process of directional information aggregation, let Ma1=100, Ma2=10, Ka1=3, Ka2=2, and the ninth preset period Ta=200ms to 2s, select 10,000 of the input side attention regulation neurons 230 of any one of the composite feature coding modules 2 as the starting neurons, select 10,000 of the concrete feature coding neurons 210 of the composite feature coding module 2 as the source neurons, and select 10,000 of the abstract feature coding neurons 220 of the composite feature coding module 2 as the target neurons.
[0280] Each of the target neurons represents an abstract, homonymous or concrete representation of the representations of the source neurons connected to it; the connection weight of a source neuron to each of the target neurons represents the correlation between the representation of the source neuron and the representation of each of the target neurons. The greater the weight, the greater the correlation, and vice versa.
[0281] For example, when the directed information aggregation process is embodied as a directed information abstraction process, the source neuron represents concrete information (such as a subclass or instance), and the target neuron represents abstract information (such as a parent class); each target neuron represents the cluster center of each source neuron connected to it (the former represents the common information component in the latter); the connection weight of a source neuron to each target neuron represents the correlation (or the distance of representation) between the information represented by the source neuron and each target neuron (i.e., the cluster center), and the larger the weight, the higher the correlation (i.e., the closer the distance of representation); this directed information abstraction process is also a clustering process, which is also a meta-learning process;
[0282] If the current target neuron is used as a new source neuron, and another group of neurons is selected as a new target neuron, the directed information aggregation process is executed, and this process is repeated iteratively, a higher level of abstract information representation can be continuously formed.
[0283] In this embodiment, the information transcription process is as follows:
[0284] Step f1: selecting a number of the neurons as starter neurons;
[0285] Step f2: selecting a number of direct downstream neurons or indirect downstream neurons of the initiating neuron as source neurons;
[0286] Step f3: selecting a number of direct downstream neurons or indirect downstream neurons of the initiating neuron as target neurons;
[0287] Step f4: enabling each of the oscillating neurons to generate a firing distribution and remain activated for a seventh preset period Tj;
[0288] Step f5: in the seventh preset period Tj, a number of the source neurons are activated;
[0289] Step f6: In the seventh preset period Tj, if a certain oscillating neuron is a direct upstream neuron of a certain target neuron, the weight of the unidirectional or bidirectional excitatory / inhibitory connection between the two is adjusted through the synaptic plasticity process; if a certain oscillating neuron is an indirect upstream neuron of a certain target neuron, the weight of the unidirectional or bidirectional excitatory / inhibitory connection between the direct upstream neuron of the target neuron and the target neuron in the connection path between the two is adjusted through the synaptic plasticity process;
[0290] Step f7: In the seventh preset period Tj, each of the target neurons may establish connections with a plurality of other target neurons, and the weights may be adjusted through the synaptic plasticity process;
[0291] Step f8: In the seventh preset period Tj, if there is a unidirectional or bidirectional excitatory connection between a source neuron and a target neuron, the weight can be adjusted through the synaptic plasticity process.
[0292] For example, in the information transcription process, the seventh preset period Tj is configured to be 20ms to 500ms, and 10,000 of the input-side attention regulation neurons 230 of any one of the composite feature coding modules 2 are selected as starter neurons, and 10,000 of the concrete feature coding neurons 210 of the composite feature coding module 2 are selected as source neurons, and 10,000 of the abstract feature coding neurons 220 of the composite feature coding module 2 are selected as target neurons.
[0293] In the process of information transcription, the information represented by part or all of the input connection weights of each activated source neuron is approximately coupled into part or all of the input connection weights of each target neuron, that is, the information is transcribed from the former to the latter; the "approximate coupling" is because the transcribed information component is also coupled with the emission distribution of each of the activating neurons, as well as the connection pathways between the activating neurons and the activated source neurons, and the connection and emission conditions of each neuron in the connection pathways between the activating neurons and the target neurons.
[0294] Specifically, in the information transcription process, if some activated activating neurons are the direct upstream neurons of some activated source neurons and some target neurons, then the connection weights between these activating neurons and these source neurons will be almost proportionally superimposed on the connection weights between these activating neurons and these target neurons, eventually making the latter approach the former; conversely, if some activated activating neurons are the indirect upstream neurons of some activated source neurons or some target neurons, then the connection weights between these activating neurons and these target neurons will eventually include the connection pathways between the activating neurons and the activated source neurons, and the influence of the connection and firing conditions of each neuron in the connection pathways between the activating neurons and the target neurons.
[0295] In this embodiment, the memory forgetting process includes an upstream issuance-dependent memory forgetting process, a downstream issuance-dependent memory forgetting process, and an upstream and downstream issuance-dependent memory forgetting process;
[0296] The upstream firing-dependent memory forgetting process is as follows: for a certain connection, if its upstream neuron continues to not fire within a fourth preset period (such as 20 minutes to 24 hours), the absolute value of the weight decreases, and the decrease is recorded as DwDecay1;
[0297] The downstream firing-dependent memory forgetting process is as follows: for a certain connection, if its downstream neuron continues to not fire within the fifth preset period (such as 20 minutes to 24 hours), the absolute value of the weight decreases, and the decrease is recorded as DwDecay2;
[0298] The upstream and downstream firing-dependent memory forgetting process is as follows: for a certain connection, if the upstream and downstream neurons do not fire synchronously for a sixth preset period (such as 20 minutes to 24 hours), the absolute value of the weight decreases, and the decrease is recorded as DwDecay3.
[0299] The synchronous release is: when the downstream neuron involved in the connection is excited, and the time interval from the current or past most recent upstream neuron excitation does not exceed the fourth preset time interval Te1, or when the upstream neuron involved in the connection is excited, and the time interval from the current or past most recent downstream neuron excitation does not exceed the fifth preset time interval Te2.
[0300] For example, let the fourth preset time interval Te1 = 30 ms, and the fifth preset time interval Te2 = 20 ms;
[0301] In the memory forgetting process, if a connection has a specified lower limit for the absolute value of the weight, the absolute value of the weight will no longer decrease when it reaches the lower limit, or the connection will be pruned.
[0302] In this embodiment, DwDecay1, DwDecay2, and DwDecay3 are respectively proportional to the weights of the involved connections, for example, DwDecay1=Kdecay1*weight, DwDecay2=Kdecay2*weight, and DwDecay1=Kdecay3*weight; let Kdecay1=Kdecay2=Kdecay3=0.01, and weight is the connection weight.
[0303] In this embodiment, the memory self-consolidation process is as follows: when a certain neuron is self-excited, the weights of some or all of the input connections of the neuron are adjusted through the unipolar downstream release-dependent synaptic enhancement process and the unipolar downstream pulse-dependent synaptic enhancement process, and the weights of some or all of the output connections of the neuron are adjusted through the unipolar upstream release-dependent synaptic enhancement process and the unipolar upstream pulse-dependent synaptic enhancement process.
[0304] The memory self-consolidation process helps to maintain the encoding of some of these neurons with near-fidelity, avoiding forgetting.
[0305] In this embodiment, the information component adjustment process is:
[0306] Step i1: Selecting a number of the neurons as starter neurons;
[0307] Step i2: selecting a number of direct downstream neurons or indirect downstream neurons of the initiating neuron as target neurons;
[0308] Step i3: enabling each of the oscillating neurons to generate a firing distribution and keep it activated within a first preset period Tb;
[0309] Step i4: In the first preset period Tb, Mb1 target neurons are activated, among which the first Kb1 target neurons with the highest activation intensity or the largest firing rate or the first firing are recorded as Gb1, and the remaining Mb1-Kb1 activated target neurons are recorded as Gb2;
[0310] Step i5: if a certain oscillating neuron is a direct upstream neuron of a certain target neuron in the Gb1, the unidirectional or bidirectional connection between the two is subjected to one or more synaptic weight enhancement processes; if a certain oscillating neuron is an indirect upstream neuron of a certain target neuron in the Gb1, the unidirectional or bidirectional connection between the direct upstream neuron of the target neuron and the target neuron in the connection path between the two is subjected to one or more synaptic weight enhancement processes;
[0311] Step i6: if a certain oscillating neuron is a direct upstream neuron of a certain target neuron in the Gb2, the unidirectional or bidirectional connection between the two is subjected to one or more synaptic weight weakening processes; if a certain oscillating neuron is an indirect upstream neuron of a certain target neuron in the Gb2, the unidirectional or bidirectional connection between the direct upstream neuron of the target neuron and the target neuron in the connection path between the two is subjected to one or more synaptic weight weakening processes;
[0312] Step i7: Each execution of the process from step i1 to step i6 is recorded as one iteration, and one or more iterations are performed;
[0313] In step i5 and step i6, after performing one or more synaptic weight enhancement processes or synaptic weight reduction processes, the weights of some or all input connections of each target neuron may be normalized or not normalized;
[0314] A number of said target neurons may be mapped to corresponding labels as a result of said information component adjustment process;
[0315] The synaptic weight enhancement process may adopt the unipolar upstream and downstream release-dependent synaptic enhancement process, or the unipolar pulse time-dependent synaptic enhancement process;
[0316] The synaptic weight weakening process may adopt the unipolar upstream and downstream release-dependent synaptic weakening process, or the unipolar pulse time-dependent synaptic weakening process;
[0317] The synaptic weight enhancement process and the synaptic weight reduction process may also respectively adopt the asymmetric bipolar pulse time-dependent synaptic plasticity process or the symmetric bipolar pulse time-dependent synaptic plasticity process.
[0318] When Kb1 takes a smaller value (for example, 1), only the target neuron with the highest activation intensity or the largest firing rate or the first to fire will undergo the synaptic weight enhancement process, that is, the information components represented by the current firing of each of the said starting neurons are superimposed to a certain extent, so that the target neuron consolidates its existing representation; while the other target neurons will undergo the synaptic weight weakening process, that is, the information components represented by the current firing of each of the said starting neurons are subtracted (decoupled) to a certain extent; therefore, multiple iterations are performed, and each iteration causes each of the said starting neurons to produce a different firing distribution, so that the representations of each of the said target neurons can be decoupled from each other; if multiple iterations are further performed to strengthen the decoupling, the representations of each of the said target neurons will become a set of relatively independent bases in the representation space.
[0319] Similarly, when Kb1 takes a larger value (for example, 8), multiple iterations are performed, and each iteration causes each of the excitatory neurons to produce a different firing distribution, which can make the information components represented by multiple target neurons overlap with each other to a certain extent. If multiple iterations are further performed, the representations of multiple target neurons can be close to each other.
[0320] Therefore, adjusting the Kb1 can adjust the information components represented by each target neuron.
[0321] For example, in the process of adjusting the information component, the first preset period Tb is configured to be 100ms to 500ms, and 10,000 of the input-side attention regulation neurons 230 of any one of the composite feature coding modules 2 are selected as the starting neurons, and 10,000 of the concrete feature coding neurons 210 of the composite feature coding module 2 are selected as the target neurons.
[0322] In this embodiment, the reinforcement learning process is: when a number of the connections receive reinforcement signals, in a second preset time interval, the weights of these connections change, or the weight reduction of these connections in the memory forgetting process changes, or the weight increase / weight reduction of these connections in the synaptic plasticity process changes; or,
[0323] When several of the neurons receive a reinforcement signal, within a third preset time interval (for example, within 30 seconds from the receipt of the reinforcement signal), these neurons receive positive or negative inputs, or the weights of some or all of the input connections or output connections of these neurons change, or the weight reduction amount of these connections in the memory forgetting process changes, or the weight increase / weight reduction amount of these connections in the synaptic plasticity process changes.
[0324] In this embodiment, the reinforcement signal is a constant value when the neural network has no input information; in the supervised learning process, if the result of the memory triggering process is correct, the reinforcement signal rises, and if the result of the memory triggering process is wrong, the reinforcement signal falls.
[0325] For example, the constant value of the reinforcement signal is 0. If the supervised learning process is performed, the result of the memory triggering process is correct, the reinforcement signal rises to +10, and the bidirectional excitatory connections between several concrete feature encoding neurons 210 receive the reinforcement signal (+10). In the second preset time interval (within 30 seconds from the receipt of the reinforcement signal), if these connections undergo the symmetric bipolar pulse time-dependent synaptic plasticity process, the DwLTP6 is increased by 10 based on its original value.
[0326] In this embodiment, the novelty signal modulation process is: when several of the neurons receive the novelty signal, within a sixth preset time interval (for example, within 30 seconds from the receipt of the novelty signal), these neurons receive positive or negative inputs, or the weights of some or all of the input connections or output connections of these neurons change, or the weight reduction amount of these connections in the memory forgetting process changes, or the weight increase amount / weight reduction amount of these connections in the synaptic plasticity process changes.
[0327] In this embodiment, the novelty signal is a constant value or gradually weakens over time when the neural network has no input information; the novelty signal is negatively correlated with the activation intensity or firing rate of each neuron in the target area in the memory triggering process when the neural network has input information.
[0328] For example, when there is no input information, the novelty signal is a constant value of +50; when input information (pictures or video streams) is applied, if these input information do not trigger memory information with sufficient relevance (for example, the highest activation intensity of neurons in the target area when the current picture is input is only 10% of the highest activation intensity when the picture that has been formed into memory encoding is input, then the relevance is only 10%), then the novelty signal rises from the constant value of +50 to +90;
[0329] When a number of the neurons constituting the forward neural pathway receive a novelty signal of +90, these neurons receive a positive input (e.g., +40) during a sixth preset time interval;
[0330] When a number of the neurons constituting the reverse neural pathway receive a novelty signal of +90, these neurons receive a negative input (e.g., -40) during a sixth preset time interval;
[0331] When the input information remains unchanged and the forward learning process is carried out, if the highest activation intensity of neurons in the target area is 90%, the novelty signal drops from +90 to +10;
[0332] When a number of the neurons constituting the forward neural pathway receive a novelty signal of +10, these neurons receive a negative input (e.g., -40) during a sixth preset time interval;
[0333] When a number of the neurons constituting the reverse neural pathway receive a novelty signal of +10, these neurons receive a positive input (e.g., +40) during a sixth preset time interval;
[0334] Therefore, when sufficiently novel external input information appears, the novelty signal will cause the neurons in the forward neural pathway to receive positive input (increased excitability) and thus be easier to activate, while causing the neurons in the reverse neural pathway to receive negative input (decreased excitability) and thus be more difficult to activate, thereby causing the neural network to preferentially pay attention to, identify, and learn the current novel external input information through the bottom-up forward neural pathway. Conversely, when the external input information is not novel enough, the novelty signal will cause the neurons in the forward neural pathway to receive negative input (decreased excitability) and become more difficult to activate, while causing the neurons in the reverse neural pathway to receive positive input (increased excitability) and become easier to activate, thereby causing the neural network to preferentially trigger existing memory information through the top-down reverse neural pathway, or to perform a pattern completion process, an association process, or an imagination process.
[0335] In this embodiment, the supervised learning process is:
[0336] Step r1: Given the positive firing distribution range of each neuron in the target area, and also given the negative firing distribution range of each neuron in the target area, execute step r2;
[0337] Step r2: Perform the memory triggering process. If the actual firing distribution of each neuron in the target area does not conform to the positive firing distribution range or the negative firing distribution range, it is deemed that each neuron in the target area does not encode relevant memory information, and step r3 is executed; if the actual firing distribution of each neuron in the target area conforms to the positive firing distribution range, the result of the memory triggering process is deemed to be correct, and the supervised learning process is terminated; if the actual firing distribution of each neuron in the target area conforms to the negative firing distribution range, the result of the memory triggering process is deemed to be wrong, and step r3 is executed;
[0338] Step r3: performing any one or several of the novelty signal modulation process, reinforcement learning process, active attention process, automatic attention process, directional start process, forward learning process, information aggregation process, directional information aggregation process, information component adjustment process, information transcription process, and differential information decoupling process, so that each neuron in the target area encodes relevant memory information, and executing step r1;
[0339] The supervised learning process can also be:
[0340] Step q1: Given a positive label range, and also a negative label range, execute step q2;
[0341] Step q2: Perform the memory triggering process, map the actual firing distribution of each neuron in the target area to the corresponding label. If the corresponding label does not meet the positive label range or the negative label range, it is deemed that each neuron in the target area does not encode relevant memory information, and execute step q3; if the corresponding label meets the positive label range, the result of the memory triggering process is deemed to be correct, and end this supervised learning process; if the corresponding label meets the negative label range, the result of the memory triggering process is deemed to be wrong, and execute step q3;
[0342] Step q3: perform any one or several of the novelty signal modulation process, reinforcement learning process, active attention process, automatic attention process, directional start process, forward learning process, information aggregation process, directional information aggregation process, information component adjustment process, information transcription process, and differential information decoupling process, so that each neuron in the target area encodes relevant memory information, and execute step q1.
[0343] In this embodiment, the unipolar upstream firing-dependent synaptic plasticity process includes a unipolar upstream firing-dependent synaptic enhancement process and a unipolar upstream firing-dependent synaptic weakening process;
[0344] The unipolar upstream firing-dependent synaptic enhancement process is: when the activation intensity or firing rate of the upstream neuron involved in the connection is not zero, the absolute value of the connection weight increases, and the increase is recorded as DwLTP1u;
[0345] The unipolar upstream firing-dependent synaptic weakening process is: when the activation intensity or firing rate of the upstream neuron involved in the connection is not zero, the absolute value of the connection weight decreases, and the decrease is recorded as DwLTD1u;
[0346] The DwLTP1u and DwLTD1u are non-negative values.
[0347] In this embodiment, the values of DwLTP1u and DwLTD1u in the unipolar upstream firing-dependent synaptic plasticity process include any one or more of the following:
[0348] The DwLTP1u and DwLTD1u are non-negative values, which are respectively proportional to the activation intensity or firing rate of the upstream neurons involved in the connection; or,
[0349] The DwLTP1u and DwLTD1u are non-negative values, which are respectively proportional to the activation strength or firing rate of the upstream neurons of the connection involved and the weight of the connection involved.
[0350] For example, let DwLTP1u=0.01*Fru1, DwLTD1u=0.01*Fru1, where Fru1 is the firing rate of the upstream neuron.
[0351] In this embodiment, the unipolar downstream firing-dependent synaptic plasticity process includes a unipolar downstream firing-dependent synaptic enhancement process and a unipolar downstream firing-dependent synaptic weakening process;
[0352] The unipolar downstream firing-dependent synaptic enhancement process is: when the activation intensity or firing rate of the downstream neuron involved in the connection is not zero, the absolute value of the connection weight increases, and the increase is recorded as DwLTP1d;
[0353] The unipolar downstream firing-dependent synaptic weakening process is: when the activation intensity or firing rate of the downstream neuron involved in the connection is not zero, the absolute value of the connection weight decreases, and the decrease is recorded as DwLTD1d;
[0354] The DwLTP1d and DwLTD1d are non-negative values.
[0355] In this embodiment, the values of DwLTP1d and DwLTD1d in the unipolar downstream firing-dependent synaptic plasticity process include any one or more of the following:
[0356] The DwLTP1d and DwLTD1d are non-negative values, which are respectively proportional to the activation intensity or firing rate of the downstream neurons involved in the connection; or,
[0357] The DwLTP1d and DwLTD1d are non-negative values, which are respectively proportional to the activation strength or firing rate of the downstream neurons of the connection involved and the weight of the connection involved.
[0358] For example, let DwLTP1d=0.01*Frd1, DwLTD1d=0.01*Frd1, where Frd1 is the firing rate of the downstream neuron.
[0359] In this embodiment, the monopolar upstream and downstream release-dependent synaptic plasticity process includes a monopolar upstream and downstream release-dependent synaptic enhancement process and a monopolar upstream and downstream release-dependent synaptic weakening process;
[0360] The unipolar upstream and downstream firing-dependent synaptic enhancement process is: when the activation intensity or firing rate of the upstream and downstream neurons involved in the connection is not zero, the absolute value of the connection weight increases, and the increase is recorded as DwLTP2;
[0361] The unipolar upstream and downstream firing-dependent synaptic weakening process is as follows: when the activation intensity or firing rate of the upstream neuron and the downstream neuron involved in the connection is not zero, the absolute value of the connection weight decreases, and the decrease is recorded as DwLTD2;
[0362] The DwLTP2 and DwLTD2 are non-negative values.
[0363] In this embodiment, the values of DwLTP2 and DwLTD2 in the unipolar upstream and downstream release-dependent synaptic plasticity process include any one or more of the following:
[0364] The DwLTP2 and DwLTD2 are non-negative values, which are respectively proportional to the activation intensity or firing rate of the upstream neuron involved in the connection and the activation intensity or firing rate of the downstream neuron; or,
[0365] The DwLTP2 and DwLTD2 are non-negative values, which are respectively proportional to the activation strength or firing rate of the upstream neurons of the connection involved, the activation strength or firing rate of the downstream neurons, and the weight of the connection involved.
[0366] For example, let DwLTP2 = 0.01*Fru2*Frd2, DwLTD2 = 0.01*Fru2*Frd2, Fru2 and Frd2 are the firing rates of the upstream and downstream neurons respectively.
[0367] In this embodiment, the unipolar upstream pulse-dependent synaptic plasticity process includes a unipolar upstream pulse-dependent synaptic enhancement process and a unipolar upstream pulse-dependent synaptic weakening process;
[0368] The unipolar upstream pulse-dependent synaptic enhancement process is: when the upstream neuron involved in the connection is excited, the absolute value of the connection weight increases, and the increase is recorded as DwLTP3u;
[0369] The unipolar upstream pulse-dependent synaptic weakening process is: when the upstream neuron involved in the connection is excited, the absolute value of the connection weight decreases, and the decrease is recorded as DwLTD3u;
[0370] The DwLTP3u and DwLTD3u are non-negative values.
[0371] In this embodiment, the values of DwLTP3u and DwLTD3u in the unipolar upstream pulse-dependent synaptic plasticity process include any one or more of the following:
[0372] The DwLTP3u and DwLTD3u use non-negative constants; or,
[0373] The DwLTP3u and DwLTD3u are non-negative values, which are proportional to the weights of the involved connections.
[0374] For example, let DwLTP3u=0.01*weight, DwLTD3u=0.01*weight, where weight is the connection weight.
[0375] In this embodiment, the unipolar downstream pulse-dependent synaptic plasticity process includes a unipolar downstream pulse-dependent synaptic enhancement process and a unipolar downstream pulse-dependent synaptic weakening process;
[0376] The unipolar downstream pulse-dependent synaptic enhancement process is: when the downstream neuron involved in the connection is excited, the absolute value of the connection weight increases, and the increase is recorded as DwLTP3d;
[0377] The unipolar downstream pulse-dependent synaptic weakening process is as follows: when the downstream neuron involved in the connection is excited, the absolute value of the connection weight decreases, and the decrease is recorded as DwLTD3d;
[0378] The DwLTP3d and DwLTD3d are non-negative values.
[0379] In this embodiment, the values of DwLTP3d and DwLTD3d of the unipolar downstream pulse-dependent synaptic plasticity process include any one or more of the following:
[0380] The DwLTP3d and DwLTD3d are non-negative constants; or,
[0381] The DwLTP3d and DwLTD3d are non-negative values, which are proportional to the weights of the involved connections.
[0382] For example, let DwLTP3d=0.01*weight, DwLTD3d=0.01*weight, where weight is the connection weight.
[0383] In this embodiment, the monopolar pulse time-dependent synaptic plasticity process includes a monopolar pulse time-dependent synaptic enhancement process and a monopolar pulse time-dependent synaptic weakening process;
[0384] The unipolar pulse time-dependent synaptic enhancement process is: when the downstream neuron involved in the connection is excited, and the time interval from the current or past most recent upstream neuron excitation is no more than Tg1, or when the upstream neuron involved in the connection is excited, and the time interval from the current or past most recent downstream neuron excitation is no more than Tg2, then the absolute value of the connection weight increases, and the increase is recorded as DwLTP4;
[0385] The unipolar pulse time-dependent synaptic weakening process is: when the downstream neuron involved in the connection is excited, and the time interval from the current or past last upstream neuron excitation is no more than Tg3, or when the upstream neuron involved in the connection is excited, and the time interval from the current or past last downstream neuron excitation is no more than Tg4, then the absolute value of the connection weight is reduced, and the reduction is recorded as DwLTD4;
[0386] The DwLTP4, DwLTD4, Tg1, Tg2, Tg3, and Tg4 are all non-negative values. For example, Tg1, Tg2, Tg3, and Tg4 are set to 200 ms.
[0387] In this embodiment, the values of DwLTP4 and DwLTD4 in the unipolar pulse time-dependent synaptic plasticity process include any one or more of the following:
[0388] The DwLTP4 and DwLTD4 use non-negative constants; or,
[0389] The DwLTP4 and DwLTD4 are non-negative values, which are respectively proportional to the weights of the involved connections.
[0390] For example, let DwLTP4=KLTP4*weight+C1, DwLTD4=KLTD4*weight+C2; where KLTP4=0.01, is the proportional coefficient of the synaptic enhancement process, KLTD4=0.01, is the proportional coefficient of the synaptic weakening process, C1 and C2 are constants, set to 0.001.
[0391] In this embodiment, the asymmetric bipolar pulse time-dependent synaptic plasticity process is:
[0392] When the downstream neuron involved in the connection is excited, if the time interval from the current or past last upstream neuron excitation does not exceed Th1, the absolute value of the connection weight increases, and the increase is recorded as DwLTP5; if the time interval from the current or past last upstream neuron excitation exceeds Th1 but does not exceed Th2, the absolute value of the connection weight decreases, and the decrease is recorded as DwLTD5; or,
[0393] When the upstream neuron involved in the connection is excited, if the time interval from the current or past last downstream neuron excitation does not exceed Th3, the absolute value of the connection weight increases, and the increase is recorded as DwLTP5; if the time interval from the current or past last downstream neuron excitation exceeds Th3 but does not exceed Th4, the absolute value of the connection weight decreases, and the decrease is recorded as DwLTD5;
[0394] The Th1, Th3, DwLTP5, and DwLTD5 are non-negative values, Th2 is a value greater than Th1, and Th4 is a value greater than Th3; for example, let Th1=Th3=150ms, Th2=Th4=200ms.
[0395] In this embodiment, the values of DwLTP5 and DwLTD5 in the asymmetric bipolar pulse time-dependent synaptic plasticity process include any one or more of the following:
[0396] The DwLTP5 and DwLTD5 adopt non-negative constants; or,
[0397] The DwLTP5 and DwLTD5 are non-negative values, which are respectively proportional to the weights of the involved connections. For example, let DwLTP5 = KLTP5*weight, DwLTD5 = KLTD5*weight, such as let KLTP5 = 0.01, KLTD5 = 0.01; or,
[0398] The DwLTP5 and DwLTD5 are non-negative values, DwLTP5 is negatively correlated with the time interval between the downstream neuron and the upstream neuron, when the time interval is 0, DwLTP5 reaches the specified maximum value DwLTPmax5, when the time interval is Th1, DwLTP5 is 0; DwLTD5 is negatively correlated with the time interval between the downstream neuron and the upstream neuron, when the time interval is Th1, DwLTD5 reaches the specified maximum value DwLTDmax5, when the time interval is Th2, DwLTD5 is 0; for example , let DwLTPmax5=0.1, DwLTDmax5=0.1, let DwLTP5=-DwLTPmax5 / Th1*DeltaT1+DwLTPmax5, let DwLTD5=-DwLTDmax5 / (Th2-Th1)*DeltaT1+DwLTDmax5*Th2 / (Th2-Th1), DeltaT1 is the time interval between the firing of downstream neurons and upstream neurons (that is, the time when the downstream neuron fires minus the time when the upstream neuron fires).
[0399] In this embodiment, the symmetric bipolar pulse time-dependent synaptic plasticity process is:
[0400] When the downstream neuron of the connection involved is excited, if the time interval from the current or past most recent upstream neuron excitation does not exceed Ti1, the absolute value of the connection weight increases, and the increase is recorded as DwLTP6;
[0401] When the upstream neuron of the connection involved is excited, if the time interval from the most recent downstream neuron excitation in the past does not exceed Ti2, the absolute value of the connection weight decreases, and the decrease is recorded as DwLTD6;
[0402] The Ti1, Ti2, DwLTP6, and DwLTD6 are non-negative values; for example, let Ti1 = 200 ms, Ti2 = 200 ms.
[0403] In this embodiment, the values of DwLTP6 and DwLTD6 in the symmetric bipolar pulse time-dependent synaptic plasticity process include any one or more of the following:
[0404] The DwLTP6 and DwLTD6 adopt non-negative constants; or,
[0405] The DwLTP6 and DwLTD6 are non-negative values, respectively proportional to the weights of the involved connections; for example, let DwLTP6 = KLTP6*weight, DwLTD6 = KLTD6*weight; or,
[0406] The DwLTP6 and DwLTD6 are non-negative values, and DwLTP6 is negatively correlated with the time interval between the downstream neuron and the upstream neuron. Specifically, when the time interval is 0, DwLTP6 reaches the specified maximum value DwLTPmax6, and when the time interval is Ti1, DwLTP6 is 0; DwLTD6 is negatively correlated with the time interval between the upstream neuron and the downstream neuron. When the time interval tends to 0, DwLTD6 reaches the specified maximum value DwLTDmax6, and when the time interval is Ti2, DwLTD6 is 0; for example, let DwLTPmax6=0.1, DwLTDmax6=0.1, let DwLTP6=-DwLTPmax6 / DeltaT2+DwLTPmax6, DwLTD6=-DwLTDmax6 / DeltaT3+DwLTDmax6, DeltaT2 is the time interval between the downstream neuron and the upstream neuron, and DeltaT3 is the time interval between the upstream neuron and the downstream neuron.
[0407] The beneficial effects of the present invention are as follows: the present invention discloses a brain-like visual neural network with forward learning and meta-learning functions, including a primary feature encoding module and a composite feature encoding module, including active and automatic attention mechanisms, a neural circuit with explicit encoding of position information of visual features, forward and reverse neural pathways, supporting up and down bidirectional information processing processes, and adopting a variety of biologically reasonable plasticity processes. It can perform forward learning, quickly encode visual representation information in input images or video streams into memory information, and perform information abstraction and information component modulation processes to obtain common feature information and differential feature information between objects, forming information channels with multiple information dimensions and information abstraction degrees, while improving generalization capabilities and retaining detail information. It also supports reinforcement learning, supervised learning, and novelty signal modulation processes, and does not rely on an end-to-end training paradigm of error backpropagation and gradient descent, breaking through the bottleneck of the existing deep learning theoretical system and providing a basis for the design and application of neuromorphic chips.
[0408] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0409] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A brain-like visual neural network with forward learning and meta-learning functions, characterized in that: include: Several primary feature coding modules and several composite feature coding modules; Each module includes multiple neurons; The neurons include primary feature encoding neurons, concrete feature encoding neurons, and abstract feature encoding neurons; The primary feature encoding module includes a plurality of primary feature encoding neurons, encoding primary visual feature information; The composite feature coding module includes a concrete feature coding unit and an abstract feature coding unit; The concrete feature encoding unit includes a plurality of the concrete feature encoding neurons, encoding concrete visual feature information; The abstract feature encoding unit includes a plurality of abstract feature encoding neurons, encoding abstract visual feature information; Several of the primary feature encoding neurons respectively form unidirectional or bidirectional excitatory / inhibitory connections with other several of the primary feature encoding neurons; A plurality of the primary feature coding neurons respectively form unidirectional or bidirectional excitatory / inhibitory connections with a plurality of the concrete feature coding neurons or a plurality of the abstract feature coding neurons located in at least one of the composite feature coding modules; A plurality of the concrete feature coding neurons located in the same composite feature coding module respectively form unidirectional or bidirectional excitatory / inhibitory connections with a plurality of the abstract feature coding neurons located in the same composite feature coding module; Several of the concrete feature coding neurons and the abstract feature coding neurons in several of the composite feature coding modules respectively form unidirectional or bidirectional excitatory / inhibitory connections with several of the concrete feature coding neurons and the abstract feature coding neurons in other several of the composite feature coding modules; The neural network caches and encodes information through the firing of the neurons, and encodes, stores and transmits information through the connections between the neurons; Input a picture or a video stream, multiply a plurality of pixel values of a plurality of pixels of each frame of the picture by weights and input them into a plurality of the primary feature coding neurons, so that the plurality of the primary feature coding neurons are activated; For a number of the neurons, their membrane potentials are calculated to determine whether to fire, and if so, the membrane potentials of the downstream neurons are accumulated to determine whether to fire, so that the firing is propagated in the neural network; the weight of the connection between the upstream neuron and the downstream neuron is a constant value, or is dynamically adjusted through the synaptic plasticity process; The working process of the neural network includes: forward memory process, memory triggering process, information aggregation process, directional information aggregation process, information transcription process, memory forgetting process, memory self-consolidation process, information component adjustment process, reinforcement learning process, novelty signal modulation process, and supervised learning process; The synaptic plasticity process includes a monopolar upstream release-dependent synaptic plasticity process, a monopolar downstream release-dependent synaptic plasticity process, a monopolar upstream and downstream release-dependent synaptic plasticity process, a monopolar upstream pulse-dependent synaptic plasticity process, a monopolar downstream pulse-dependent synaptic plasticity process, a monopolar pulse time-dependent synaptic plasticity process, an asymmetric bipolar pulse time-dependent synaptic plasticity process, and a symmetric bipolar pulse time-dependent synaptic plasticity process; Mapping a number of the neurons to corresponding labels as output.
2. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: Several neurons of the neural network are pulse neurons or non-pulse neurons.
3. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: Several neurons of the neural network are self-exciting neurons; the self-exciting neurons include conditional self-exciting neurons and unconditional self-exciting neurons; If the conditionally self-exciting neuron is not excited by external input in the first preset time interval, it will self-excite according to probability P; The membrane potential of the unconditionally self-exciting neuron is automatically accumulated gradually in the absence of external input. When the membrane potential reaches a threshold, the unconditionally self-exciting neuron is excited and the membrane potential is restored to the resting potential to restart the accumulation process.
4. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: Several connections of the neural network are replaced by convolution operations.
5. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The composite feature encoding module also includes an input-side attention control unit and an output-side attention control unit; The neurons also include input-side attention regulation neurons and output-side attention regulation neurons; The input-side attention regulation unit includes a plurality of input-side attention regulation neurons; The output-side attention regulation unit includes a plurality of output-side attention regulation neurons; A plurality of the input-side attention regulating neurons receive unidirectional or bidirectional excitatory / inhibitory connections from a plurality of the primary feature encoding neurons respectively; Each of the input-side attention regulation neurons forms a unidirectional or bidirectional excitatory connection with a plurality of the concrete feature encoding neurons or a plurality of the abstract feature encoding neurons of the composite feature encoding module; Each of the input-side attention regulating neurons receives unidirectional or bidirectional excitatory connections from a number of the concrete feature encoding neurons or a number of the abstract feature encoding neurons or a number of the output-side attention regulating neurons from other composite feature encoding modules; Alternatively, each of the input-side attention regulation neurons forms a unidirectional or bidirectional excitatory connection with a plurality of other input-side attention regulation neurons; Each of the output-side attention regulation neurons forms a unidirectional or bidirectional excitatory connection with a plurality of the concrete feature coding neurons or a plurality of the abstract feature coding neurons or a plurality of the input-side attention regulation neurons located in other composite feature coding modules; Each of the output-side attention regulating neurons receives unidirectional or bidirectional excitatory connections from a plurality of the concrete feature encoding neurons or a plurality of the abstract feature encoding neurons in the composite feature encoding module; Alternatively, each of the output-side attention regulation neurons forms a unidirectional or bidirectional excitatory connection with a plurality of other output-side attention regulation neurons; Each of the input-side attention regulation neurons has an input-side attention control terminal; each of the output-side attention regulation neurons has an output-side attention control terminal; The working process of the neural network also includes active attention process and automatic attention process; The active attention process is: adjusting the activation intensity or firing rate or pulse firing phase of each input-side attention control neuron by the strength of the attention control signal applied at the input-side attention control end, thereby controlling the information entering the corresponding concrete feature coding unit and the abstract feature coding unit, and adjusting the size and proportion of each information component; or, adjusting the activation intensity or firing rate or pulse firing phase of each output-side attention control neuron by the strength of the attention control signal applied at the output-side attention control end, thereby controlling the information output from the corresponding concrete feature coding unit and the abstract feature coding unit, and adjusting the size and proportion of each information component; The automatic attention process is: when a number of neurons connected to the input-side attention regulating neurons are activated, these input-side attention regulating neurons are more likely to be activated, so that the relevant information components are more likely to be input into the corresponding concrete feature encoding unit and the abstract feature encoding unit; Alternatively, when several neurons connected to the output-side attention regulation neurons are activated, these output-side attention regulation neurons are easier to activate, thereby making it easier for relevant information components to be output from the corresponding concrete feature encoding units and abstract feature encoding units.
6. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The neural network includes one or more information channels; The working process of the neural network also includes an automatic formation process of information channels; The automatic formation process of the information channel is: by performing any one or several of the forward memory process, memory triggering process, information aggregation process, directional information aggregation process, information transcription process, memory forgetting process, memory self-consolidation process, information component adjustment process, active attention process, and automatic attention process, the connection relationship and weights between the neurons are adjusted so that the neural network forms one or more information channels, each information channel encodes one or more information components; there are intersections between the information channels; or, by presetting initial connection relationships and initial parameters, the neural network forms one or more information channels, each information channel encodes one or more preset information components.
7. The brain-like visual neural network with forward learning and meta-learning functions according to claim 6, characterized in that: The information channel includes a primary information channel; The primary information channel is: all the primary feature encoding neurons and their connections constitute the primary information channel; The primary information channels include a primary contrast information channel, a primary orientation information channel, a primary edge information channel, and a primary color block information channel; The primary contrast information channel is: select a number of adjacent pixels in the input image as central area pixels, select a number of pixels around the central area pixels as peripheral area pixels, multiply a number of pixel values of each of the central area pixels and the peripheral area pixels by weights and input them into a number of primary feature encoding neurons, that is, forming a central-peripheral topological structure, and these primary feature encoding neurons and their connections constitute one or more primary contrast information channels; In the primary information channel, one or more numbers of pixels, positions and areas of adjacent pixels covering the screen space are selected, and one or more pixel values of these pixels are multiplied by one or more weights respectively to form a plurality of primary orientation information channels, primary edge information channels, primary color block information channels or a combination thereof with one or more receptive fields; The primary contrast information channel includes a light-dark contrast information channel, a dark-light contrast information channel, a red-green contrast information channel, a green-red contrast information channel, a yellow-blue contrast information channel, and a blue-yellow contrast information channel; The light-dark contrast information channel is: the R, G, B pixel values of each of the central area pixels are multiplied by positive weights and input into a plurality of the primary feature coding neurons, and the R, G, B pixel values of each of the surrounding area pixels are multiplied by negative weights and input into these primary feature coding neurons, and these primary feature coding neurons and their connections constitute the light-dark contrast information channel; The dark-light contrast information channel is: the R, G, B pixel values of each of the central area pixels are multiplied by negative weights and input into a plurality of the primary feature coding neurons, and the R, G, B pixel values of each of the surrounding area pixels are multiplied by positive weights and input into these primary feature coding neurons, and these primary feature coding neurons and their connections constitute the dark-light contrast information channel; The red-green contrast information channel is: the R, G, B pixel values of each of the central area pixels are multiplied by positive weights, negative weights, and positive weights, respectively, and input into a plurality of the primary feature coding neurons; the R, G, B pixel values of each of the surrounding area pixels are multiplied by negative weights, positive weights, and positive weights, respectively, and input into these primary feature coding neurons; these primary feature coding neurons and their connections constitute the red-green contrast information channel; The green-red contrast information channel is: the R, G, B pixel values of each of the central area pixels are multiplied by negative weights, positive weights, and positive weights, respectively, and input into a plurality of the primary feature coding neurons; the R, G, B pixel values of each of the surrounding area pixels are multiplied by positive weights, negative weights, and positive weights, respectively, and input into these primary feature coding neurons; these primary feature coding neurons and their connections constitute the green-red contrast information channel; The yellow-blue contrast information channel is: the R, G, B pixel values of each of the central area pixels are multiplied by positive weights, positive weights, and negative weights, respectively, and input into a plurality of the primary feature coding neurons; the R, G, B pixel values of each of the surrounding area pixels are multiplied by negative weights, negative weights, and positive weights, respectively, and input into these primary feature coding neurons; these primary feature coding neurons and their connections constitute the yellow-blue contrast information channel; The blue-yellow contrast information channel is: the R, G, B pixel values of each of the central area pixels are multiplied by negative weights, negative weights, and positive weights, respectively, and input into a number of the primary feature coding neurons; the R, G, B pixel values of each of the surrounding area pixels are multiplied by positive weights, positive weights, and negative weights, respectively, and input into these primary feature coding neurons. These primary feature coding neurons and their connections constitute the blue-yellow contrast information channel.
8. The brain-like visual neural network with forward learning and meta-learning functions according to claim 7, characterized in that: The primary information channel also includes a primary optical flow information channel; The primary optical flow information channel is: optical flow is calculated for a number of pixels in the input image to obtain the direction and velocity values of the optical flow movement, different direction values and velocity values are combined, and the values are multiplied by weights and input into a number of primary feature encoding neurons, and these primary feature encoding neurons and their connections constitute the primary optical flow information channel; The primary visual feature information also includes optical flow information.
9. The brain-like visual neural network with forward learning and meta-learning functions according to claim 5, characterized in that: The composite feature encoding module also includes a plurality of position encoding units; The neurons also include position coding neurons; The position encoding unit includes a plurality of position encoding neurons, encoding position information; Each of the position encoding units corresponds to a plurality of subspaces in the picture space, and each subspace has an intersection; Each of the position coding neurons corresponds to each region corresponding to the position in each subspace corresponding to the position coding unit where it is located, and receives unidirectional or bidirectional excitatory connections from a number of neurons whose receptive fields are in these regions; Several of the position coding neurons respectively form unidirectional or bidirectional excitatory connections with several other position coding neurons corresponding to the same area; Alternatively, a plurality of the position coding neurons respectively form unidirectional or bidirectional excitatory connections with a plurality of the input-side attention regulation neurons / output-side attention regulation neurons / concrete feature coding neurons / abstract feature coding neurons located in the composite feature coding module where they are located; Alternatively, several of the position coding neurons respectively form unidirectional or bidirectional excitatory connections with several of the input-side attention regulation neurons / concrete feature coding neurons / abstract feature coding neurons located in other composite feature coding modules.
10. The brain-like visual neural network with forward learning and meta-learning functions according to claim 7, characterized in that: The information channel also includes an intermediate information channel; The intermediate information channel includes an intermediate position information channel; The composite feature encoding module includes an input-side attention regulation unit and an output-side attention regulation unit; The neurons also include input-side attention regulation neurons and output-side attention regulation neurons; The input-side attention regulation unit includes a plurality of input-side attention regulation neurons; The output-side attention regulation unit includes a plurality of output-side attention regulation neurons; The neurons also include position coding neurons; The position encoding unit includes a plurality of position encoding neurons, encoding position information; The intermediate position information channel is: through the automatic formation process of the information channel, or by presetting the initial connection relationship and initial parameters, the proportion of the total weight of the connections from the position coding neurons and neurons encoding the position information in the total weight of some or all input connections of several input-side attention regulation neurons / output-side attention regulation neurons / concrete feature coding neurons / abstract feature coding neurons in several composite feature coding modules reaches or exceeds a first preset ratio, and the connection weights from the position coding neurons and neurons encoding the position information are combined in multiple proportions, so that these input-side attention regulation neurons / output-side attention regulation neurons / concrete feature coding neurons / abstract feature coding neurons respectively have one or more receptive fields and respectively encode one or more position information, and together with the position coding neurons constitute the intermediate position information channel.
11. The brain-like visual neural network with forward learning and meta-learning functions according to claim 10, characterized in that: The intermediate information channel also includes an intermediate visual feature information channel; The intermediate visual feature information channel is: through the automatic formation process of the information channel, or by presetting the initial connection relationship and initial parameters, the proportion of the total connection weight of neurons from the primary information channel in the total weight of part or all of the input connections of several input-side attention control neurons / output-side attention control neurons / concrete feature coding neurons / abstract feature coding neurons in several composite feature coding modules reaches or exceeds a second preset ratio, and the connection weights of each neuron in each area and position in the corresponding picture space of the primary information channel and the intermediate information channel are combined in multiple proportions, so that these input-side attention control neurons / output-side attention control neurons / concrete feature coding neurons / abstract feature coding neurons respectively have one or more receptive fields, respectively encode one or more intermediate visual feature information, and together constitute the intermediate visual feature information channel.
12. The brain-like visual neural network with forward learning and meta-learning functions according to claim 10, characterized in that: The information channel also includes an advanced information channel; The high-level information channel is: through the automatic formation process of the information channel, or by presetting the initial connection relationship and initial parameters, the proportion of the total connection weight of neurons from the intermediate information channel in the total weight of part or all of the input connections of several input-side attention control neurons / output-side attention control neurons / concrete feature coding neurons / abstract feature coding neurons in several composite feature coding modules reaches or exceeds a third preset ratio, and the connection weights of each neuron in each area and position in the corresponding picture space of the primary information channel, the intermediate information channel, and the high-level information channel are combined in multiple proportions, so that these input-side attention control neurons / output-side attention control neurons / concrete feature coding neurons / abstract feature coding neurons respectively have one or more receptive fields, respectively encode one or more high-level visual feature information, and together constitute the high-level information channel.
13. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The neural network includes a forward neural pathway and a reverse neural pathway; The forward neural pathway and the reverse neural pathway are respectively: a plurality of the primary feature encoding modules / composite feature encoding modules are cascaded in a first preset order, a neural pathway formed by a plurality of the neurons cascaded in the first preset order is used as the forward neural pathway, and a neural pathway formed by a plurality of the neurons cascaded in the reverse order is used as the reverse neural pathway; In each primary feature encoding module / composite feature encoding module, several neurons constituting the forward neural pathway form unidirectional or bidirectional excitatory / inhibitory connections with several neurons constituting the reverse neural pathway; The working process of the neural network also includes a directional start process, and the directional start process includes a forward start process and a reverse start process; The forward start process is: Step o1: Select several neurons in the forward neural pathway as starter neurons; Step o2: enabling each of the oscillating neurons to generate a firing distribution and maintain activation for a third preset period Tfprime; Step o3: a number of the neurons in the reverse neural pathway that receive excitatory connections from the initiating neuron receive non-negative inputs to be more easily activated; Step o4: a number of the neurons in the reverse neural pathway that receive inhibitory connections from the excitatory neurons receive non-positive inputs to become more difficult to activate; The reverse startup process is: Step n1: Select several neurons in the reverse neural pathway as starter neurons; Step n2: causing each of the oscillating neurons to generate a firing distribution and remain activated for a tenth preset period Tbprime; Step n3: a number of the neurons in the forward neural pathway that receive excitatory connections from the energizing neuron receive non-negative inputs to be more easily activated; Step n4: Some of the neurons in the forward neural pathway that receive inhibitory connections from the excitatory neuron receive non-positive inputs to make them more difficult to activate.
14. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The neurons also include interneurons; The primary feature coding module and the composite feature coding module respectively include a number of the intermediate neurons, and the several intermediate neurons respectively form unidirectional inhibitory connections with the corresponding several neurons in the corresponding modules, and the corresponding several neurons in each module respectively form unidirectional excitatory connections with the corresponding several intermediate neurons.
15. The brain-like visual neural network with forward learning and meta-learning functions according to claim 5, characterized in that: The neurons also include differential information decoupling neurons, and the working process of the neural network also includes a differential information decoupling process; The differential information decoupling process is: Selecting a number of the input-side attention regulation neurons / output-side attention regulation neurons / concrete feature encoding neurons / abstract feature encoding neurons as target neurons; Selecting a number of neurons having unidirectional / bidirectional excitatory connections with the target neuron as concrete information source neurons; Selecting several other neurons having unidirectional / bidirectional excitatory connections with the target neuron as abstract information source neurons; Each of the concrete information source neurons has a number of the differential information decoupling neurons matched with it; each of the concrete information source neurons forms a unidirectional excitatory connection with each of the matched differential information decoupling neurons; the differential information decoupling neurons form a unidirectional inhibitory connection with the target neurons or form a unidirectional inhibitory synapse-synapse connection with the connection of the information source neuron input to the target neuron, so that the signal input from the concrete information source neuron to the target neuron is subject to inhibitory regulation by the matched differential information decoupling neuron; the abstract information source neuron forms a unidirectional excitatory connection with the differential information decoupling neuron; Each of the differential information decoupling neurons has a decoupling control signal input terminal; the degree of information decoupling is adjusted by adjusting the signal size applied to the decoupling control signal input terminal.
16. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The forward learning process is: Step a1: selecting a number of the neurons as starter neurons; Step a2: selecting a number of the neurons as target neurons; Step a3: adjusting the weights of the unidirectional excitatory connections between each activated initiator neuron and a plurality of target neurons through the synaptic plasticity process; Step a4: each of the activated target neurons establishes unidirectional or bidirectional excitatory connections with a number of other target neurons, or establishes self-circulating excitatory connections with itself, and the weights of these connections are adjusted through the synaptic plasticity process; When the weights of the input / output connections of each of the target neurons are adjusted through the synaptic plasticity process, the weights of some or all of the input or output connections are normalized, or not normalized.
17. A brain-like visual neural network with forward learning and meta-learning functions according to any one of claims 1 and 3, characterized in that: The memory triggering process is: inputting information, or directly activating several neurons in the neural network, or self-exciting several neurons in the neural network, or propagating the existing activation states of several neurons in the neural network in the neural network, if it causes several neurons in the target area to fire within a second preset period, then the representation of each fired neuron in the target area together with its activation intensity or firing rate is used as the result of the memory triggering process; The target area is any sub-network in the neural network.
18. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The information aggregation process is as follows: Step g1: Selecting a number of the neurons as starter neurons; Step g2: selecting a number of the neurons as source neurons; Step g3: selecting a number of the neurons as target neurons; Step g4: causing each of the oscillating neurons to generate a firing distribution and remain activated for an eighth preset period Tk; Step g5: in the eighth preset period Tk, the weights of the unidirectional or bidirectional excitatory / inhibitory connections between each activated starting neuron and a plurality of target neurons are adjusted through the synaptic plasticity process; Step g6: In the eighth preset period Tk, the weights of the unidirectional or bidirectional excitatory / inhibitory connections between each activated source neuron and a plurality of target neurons are adjusted through the synaptic plasticity process; Step g7: Each execution of steps g1 to g6 is recorded as one iteration, and one or more iterations are performed; A number of the target neurons are mapped to corresponding labels as a result of the information aggregation process.
19. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The directional information aggregation process is as follows: Step h1: Selecting a number of the neurons as starter neurons; Step h2: selecting a number of the neurons as source neurons; Step h3: selecting a number of the neurons as target neurons; Step h4: causing each of the oscillating neurons to generate a firing distribution and remain activated for a ninth preset period Ta; Step h5: in the ninth preset period Ta, Ma1 of the source neurons are activated, and Ma2 of the target neurons are activated; Step h6: In the ninth preset period Ta, the first Ka1 source neurons with the highest activation intensity or the largest firing rate or the first to start firing are recorded as Ga1, and the remaining Ma1-Ka1 activated source neurons are recorded as Ga2; Step h7: In the ninth preset period Ta, the first Ka2 target neurons with the highest activation intensity or the largest firing rate or the first firing are recorded as Ga3, and the remaining Ma2-Ka2 activated target neurons are recorded as Ga4; Step h8: In the ninth preset cycle Ta, the unidirectional or bidirectional excitatory / inhibitory connections of each source neuron in Ga1 and a plurality of target neurons in Ga3 are respectively subjected to one or more synaptic weight enhancement processes; Step h9: In the ninth preset cycle Ta, the unidirectional or bidirectional excitatory / inhibitory connections of each source neuron in Ga1 and a plurality of target neurons in Ga4 are respectively subjected to one or more synaptic weight weakening processes; Step h10: In the ninth preset period Ta, the unidirectional or bidirectional excitatory / inhibitory connections between each source neuron in Ga2 and a plurality of target neurons in Ga3 respectively undergo one or more synaptic weight weakening processes or not; Step h11: In the ninth preset cycle Ta, the unidirectional or bidirectional excitatory / inhibitory connections between each source neuron in Ga2 and a plurality of target neurons in Ga4 are respectively subjected to one or more synaptic weight enhancement processes or not; Step h12: In the ninth preset period Ta, each activated starting neuron performs one or more synaptic weight enhancement processes with the unidirectional or bidirectional excitatory / inhibitory connections of several target neurons in Ga3; Step h13: In the ninth preset period Ta, each activated starting neuron performs one or more synaptic weight weakening processes with the unidirectional or bidirectional excitatory / inhibitory connections of several target neurons in Ga4; Step h14: Each execution of the process from step h1 to step h13 is recorded as one iteration, and one or more iterations are performed; In the process of steps h8 to h13, after performing one or more synaptic weight enhancement processes or synaptic weight reduction processes, the weights of some or all input or output connections of each source neuron or target neuron are normalized, or not normalized; The synaptic weight enhancement process adopts the unipolar upstream and downstream release-dependent synaptic enhancement process, or the unipolar pulse time-dependent synaptic enhancement process; The synaptic weight weakening process adopts the unipolar upstream and downstream release-dependent synaptic weakening process, or the unipolar pulse time-dependent synaptic weakening process; The synaptic weight enhancement process and the synaptic weight weakening process also respectively adopt the asymmetric bipolar pulse time-dependent synaptic plasticity process or the symmetric bipolar pulse time-dependent synaptic plasticity process; Ma1 and Ma2 are positive integers, Ka1 is a positive integer not exceeding Ma1, and Ka2 is a positive integer not exceeding Ma2.
20. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The information transcription process is as follows: Step f1: selecting a number of the neurons as starter neurons; Step f2: selecting a number of direct downstream neurons or indirect downstream neurons of the initiating neuron as source neurons; Step f3: selecting a number of direct downstream neurons or indirect downstream neurons of the initiating neuron as target neurons; Step f4: enabling each of the oscillating neurons to generate a firing distribution and remain activated for a seventh preset period Tj; Step f5: in the seventh preset period Tj, a number of the source neurons are activated; Step f6: In the seventh preset period Tj, if a certain oscillating neuron is a direct upstream neuron of a certain target neuron, the weight of the unidirectional or bidirectional excitatory / inhibitory connection between the two is adjusted through the synaptic plasticity process; if a certain oscillating neuron is an indirect upstream neuron of a certain target neuron, the weight of the unidirectional or bidirectional excitatory / inhibitory connection between the direct upstream neuron of the target neuron and the target neuron in the connection path between the two is adjusted through the synaptic plasticity process; Step f7: in the seventh preset period Tj, each of the target neurons establishes connections with a number of other target neurons respectively, and adjusts the weights through the synaptic plasticity process; Step f8: In the seventh preset period Tj, if there is a unidirectional or bidirectional excitatory connection between a source neuron and a target neuron, the weight is adjusted through the synaptic plasticity process.
21. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The memory forgetting process includes an upstream issuance-dependent memory forgetting process, a downstream issuance-dependent memory forgetting process, and an upstream and downstream issuance-dependent memory forgetting process; The upstream firing-dependent memory forgetting process is as follows: for a certain connection, if its upstream neuron continues not firing within the fourth preset period, the absolute value of the weight decreases, and the decrease is recorded as DwDecay1; The downstream firing-dependent memory forgetting process is as follows: for a certain connection, if its downstream neuron continues to not fire within the fifth preset cycle, the absolute value of the weight decreases, and the decrease is recorded as DwDecay2; The upstream and downstream firing dependent memory forgetting process is: for a certain connection, if the upstream and downstream neurons do not fire synchronously within the sixth preset period, the absolute value of the weight decreases, and the decrease is recorded as DwDecay3; The synchronous firing is: when the downstream neuron involved in the connection is excited, and the time interval from the current or past last upstream neuron excitation is not more than the fourth preset time interval Te1, or when the upstream neuron involved in the connection is excited, and the time interval from the current or past last downstream neuron excitation is not more than the fifth preset time interval Te2; In the memory forgetting process, if a connection has a specified lower limit for the absolute value of the weight, the absolute value of the weight will no longer decrease when it reaches the lower limit, or the connection will be pruned.
22. A brain-like visual neural network with forward learning and meta-learning functions according to any one of claims 1 and 3, characterized in that: The memory self-consolidation process is as follows: when a certain neuron is self-excited, the weights of some or all of the input connections of the neuron are adjusted through the unipolar downstream release-dependent synaptic enhancement process and the unipolar downstream pulse-dependent synaptic enhancement process, and the weights of some or all of the output connections of the neuron are adjusted through the unipolar upstream release-dependent synaptic enhancement process and the unipolar upstream pulse-dependent synaptic enhancement process.
23. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The information component adjustment process is: Step i1: Select a number of the neurons as starter neurons; Step i2: selecting a number of direct downstream neurons or indirect downstream neurons of the initiating neuron as target neurons; Step 13: causing each of the oscillating neurons to generate a firing distribution and keep it activated within a first preset period Tb; Step i4: In the first preset period Tb, Mb1 target neurons are activated, among which the first Kb1 target neurons with the highest activation intensity or the largest firing rate or the first firing are recorded as Gb1, and the remaining Mb1-Kb1 activated target neurons are recorded as Gb2; Step i5: if a certain oscillating neuron is a direct upstream neuron of a certain target neuron in the Gb1, the unidirectional or bidirectional connection between the two is subjected to one or more synaptic weight enhancement processes; if a certain oscillating neuron is an indirect upstream neuron of a certain target neuron in the Gb1, the unidirectional or bidirectional connection between the direct upstream neuron of the target neuron and the target neuron in the connection path between the two is subjected to one or more synaptic weight enhancement processes; Step i6: if a certain oscillating neuron is a direct upstream neuron of a certain target neuron in the Gb2, the unidirectional or bidirectional connection between the two is subjected to one or more synaptic weight weakening processes; if a certain oscillating neuron is an indirect upstream neuron of a certain target neuron in the Gb2, the unidirectional or bidirectional connection between the direct upstream neuron of the target neuron and the target neuron in the connection path between the two is subjected to one or more synaptic weight weakening processes; Step i7: Each execution of step i1 to step i6 is recorded as one iteration, and one or more iterations are performed; In step i5 and step i6, after performing one or more synaptic weight enhancement processes or synaptic weight reduction processes, the weights of some or all input connections of each target neuron are normalized, or not normalized; The synaptic weight enhancement process adopts the unipolar upstream and downstream release-dependent synaptic enhancement process, or the unipolar pulse time-dependent synaptic enhancement process; The synaptic weight weakening process adopts the unipolar upstream and downstream release-dependent synaptic weakening process, or the unipolar pulse time-dependent synaptic weakening process; The synaptic weight enhancement process and the synaptic weight reduction process also respectively adopt the asymmetric bipolar pulse time-dependent synaptic plasticity process or the symmetric bipolar pulse time-dependent synaptic plasticity process.
24. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The reinforcement learning process is: when a number of the connections receive reinforcement signals, in a second preset time interval, the weights of these connections change, or the weight reduction of these connections in the memory forgetting process changes, or the weight increase / weight reduction of these connections in the synaptic plasticity process changes; or, When a number of the neurons receive a reinforcement signal, in a third preset time interval, the neurons receive a positive or negative input, or the weights of some or all of the input connections or output connections of the neurons change, or the weight reduction of the connections in the memory forgetting process changes, or the weight increase / weight reduction of the connections in the synaptic plasticity process changes; The reinforcement signal is a constant value when the neural network has no input information; during the supervised learning process, if the result of the memory triggering process is correct, the reinforcement signal rises, and if the result of the memory triggering process is wrong, the reinforcement signal falls.
25. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The novelty signal modulation process is: when a number of the neurons receive the novelty signal, in a sixth preset time interval, the neurons receive positive or negative inputs, or the weights of some or all of the input connections or output connections of the neurons change, or the weight reduction of the connections in the memory forgetting process changes, or the weight increase / weight reduction of the connections in the synaptic plasticity process changes; The novelty signal is a constant value or gradually weakens over time when the neural network has no input information; the novelty signal is negatively correlated with the activation intensity or firing rate of each neuron in the target area in the memory triggering process when the neural network has input information.
26. A brain-like visual neural network with forward learning and meta-learning functions according to any one of claims 1, 24 and 25, characterized in that: The supervised learning process is: Step r1: given the positive firing distribution range of each neuron in the target area, given the negative firing distribution range of each neuron in the target area, execute step r2; Step r2: Perform the memory triggering process. If the actual firing distribution of each neuron in the target area does not conform to the positive firing distribution range or the negative firing distribution range, it is deemed that each neuron in the target area does not encode relevant memory information, and step r3 is executed; if the actual firing distribution of each neuron in the target area conforms to the positive firing distribution range, the result of the memory triggering process is deemed to be correct, and the supervised learning process is terminated; if the actual firing distribution of each neuron in the target area conforms to the negative firing distribution range, the result of the memory triggering process is deemed to be wrong, and step r3 is executed; Step r3: performing any one or several of the novelty signal modulation process, reinforcement learning process, active attention process, automatic attention process, directional start process, forward learning process, information aggregation process, directional information aggregation process, information component adjustment process, information transcription process, and differential information decoupling process, so that each neuron in the target area encodes relevant memory information, and executing step r1; Alternatively, the supervised learning process is: Step q1: Given a positive label range, given a negative label range, execute step q2; Step q2: Perform the memory triggering process, map the actual firing distribution of each neuron in the target area to the corresponding label. If the corresponding label does not meet the positive label range or the negative label range, it is deemed that each neuron in the target area does not encode relevant memory information, and execute step q3; if the corresponding label meets the positive label range, the result of the memory triggering process is deemed to be correct, and end this supervised learning process; if the corresponding label meets the negative label range, the result of the memory triggering process is deemed to be wrong, and execute step q3; Step q3: perform any one or several of the novelty signal modulation process, reinforcement learning process, active attention process, automatic attention process, directional start process, forward learning process, information aggregation process, directional information aggregation process, information component adjustment process, information transcription process, and differential information decoupling process, so that each neuron in the target area encodes relevant memory information, and execute step q1.
27. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The monopolar upstream firing-dependent synaptic plasticity process includes a monopolar upstream firing-dependent synaptic enhancement process and a monopolar upstream firing-dependent synaptic weakening process; The unipolar upstream firing-dependent synaptic enhancement process is: when the activation intensity or firing rate of the upstream neuron involved in the connection is not zero, the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP1 u; The unipolar upstream firing-dependent synaptic weakening process is: when the activation intensity or firing rate of the upstream neuron involved in the connection is not zero, the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD1 u; The DwLTP1 u and DwLTD1 u are non-negative values.
28. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The monopolar downstream firing-dependent synaptic plasticity process includes a monopolar downstream firing-dependent synaptic enhancement process and a monopolar downstream firing-dependent synaptic weakening process; The unipolar downstream firing-dependent synaptic enhancement process is: when the activation intensity or firing rate of the downstream neuron involved in the connection is not zero, the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP1 d; The unipolar downstream firing-dependent synaptic weakening process is: when the activation intensity or firing rate of the downstream neuron involved in the connection is not zero, the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD1 d; The DwLTP1d and DwLTD1d are non-negative values.
29. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The monopolar upstream and downstream release-dependent synaptic plasticity process includes a monopolar upstream and downstream release-dependent synaptic enhancement process and a monopolar upstream and downstream release-dependent synaptic weakening process; The unipolar upstream and downstream firing-dependent synaptic enhancement process is: when the activation intensity or firing rate of the upstream neuron and the downstream neuron involved in the connection is not zero, the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP2; The unipolar upstream and downstream firing-dependent synaptic weakening process is: when the activation intensity or firing rate of the upstream neuron and the downstream neuron involved in the connection is not zero, the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD2; The DwLTP2 and DwLTD2 are non-negative values.
30. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The monopolar upstream pulse-dependent synaptic plasticity process includes a monopolar upstream pulse-dependent synaptic enhancement process and a monopolar upstream pulse-dependent synaptic weakening process; The unipolar upstream pulse-dependent synaptic enhancement process is: when the upstream neuron involved in the connection is excited, the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP3u; The unipolar upstream pulse-dependent synaptic weakening process is: when the upstream neuron involved in the connection is excited, the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD3u; The DwLTP3u and DwLTD3u are non-negative values.
31. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The monopolar downstream pulse-dependent synaptic plasticity process includes a monopolar downstream pulse-dependent synaptic enhancement process and a monopolar downstream pulse-dependent synaptic weakening process; The unipolar downstream pulse-dependent synaptic enhancement process is: when the downstream neuron involved in the connection is excited, the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP3d; The unipolar downstream pulse-dependent synaptic weakening process is: when the downstream neuron involved in the connection is excited, the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD3d; The DwLTP3d and DwLTD3d are non-negative values.
32. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The monopolar pulse time-dependent synaptic plasticity process includes a monopolar pulse time-dependent synaptic enhancement process and a monopolar pulse time-dependent synaptic weakening process; The unipolar pulse time-dependent synaptic enhancement process is: when the downstream neuron involved in the connection is excited, and the time interval from the current or past most recent upstream neuron excitation is no more than Tg1, or when the upstream neuron involved in the connection is excited, and the time interval from the current or past most recent downstream neuron excitation is no more than Tg2, then the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP4: The unipolar pulse time-dependent synaptic weakening process is: when the downstream neuron involved in the connection is excited, and the time interval from the current or past latest upstream neuron excitation is no more than Tg3, or when the upstream neuron involved in the connection is excited, and the time interval from the current or past latest downstream neuron excitation is no more than Tg4, then the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD4: The DwLTP4, DwLTD4, Tg1, Tg2, Tg3, and Tg4 are all non-negative values.
33. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The asymmetric bipolar pulse time-dependent synaptic plasticity process is: When the downstream neuron involved in the connection is excited, if the time interval from the current or past most recent upstream neuron excitation does not exceed Th1, the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP5; if the time interval from the current or past most recent upstream neuron excitation exceeds Th1 but does not exceed Th2, the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD5; or, When the upstream neuron involved in the connection is excited, if the time interval from the current or past most recent downstream neuron excitation does not exceed Th3, the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP5; if the time interval from the current or past most recent downstream neuron excitation exceeds Th3 but does not exceed Th4, the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD5; Th1, Th3, DwLTP5, and DwLTD5 are non-negative values, Th2 is a value greater than Th1, and Th4 is a value greater than Th3.
34. The brain-like visual neural network with forward learning and meta-learning functions according to claim 1, characterized in that: The symmetric bipolar pulse time-dependent synaptic plasticity process is: When the downstream neuron involved in the connection is excited, if the time interval from the current or past most recent upstream neuron excitation does not exceed Ti 1, the absolute value of the connection weight increases, and the increase in the absolute value of the connection weight is recorded as DwLTP6; When the upstream neuron of the connection involved is excited, if the time interval from the most recent downstream neuron excitation in the past does not exceed Ti2, the absolute value of the connection weight decreases, and the decrease in the absolute value of the connection weight is recorded as DwLTD6; The Ti 1, Ti 2, DwLTP6, and DwLTD6 are non-negative values.
