Causal characterization decision-making method based on group coding
Through a causal representation decision-making method based on group coding, using technologies such as pulse neural networks and pulse gated cycle units, the problem of difficulty in capturing decision lag in the dynamic environment in the existing technology is solved, and more efficient decision adaptability and accuracy.
Patent Information
- Application Number
- CN202411084700.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-12
AI Technical Summary
Existing decision-making systems are difficult to adapt to environmental changes when dealing with complex, dynamic and uncertain environments, and are difficult to capture dynamic causal factors, resulting in decision-making lag.
Using a causal representation decision-making method based on group coding, by obtaining target environment information, converting it into a target pulse sequence and inputting a pulse neural network, obtaining decision probability, outputting decision results, combining pulse gated cyclic unit and pulse attention layer to enhance the network Adaptability and generalization ability.
It improves adaptability and decision-making accuracy to dynamic environments, reduces model training time, and enhances the network's information processing efficiency and generalization ability.
Smart Images

Figure CN118940809A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a causal representation decision method, device, equipment and medium based on group coding. Background Art
[0002] Decision-making systems need to make appropriate responses and decisions based on complex, dynamic and uncertain environmental factors.
[0003] In related technologies, decision systems are usually based on predefined rules or models, which are difficult to adapt to dynamic changes in the environment. Classic machine learning algorithms such as support vector machines and decision trees can learn patterns from historical data, but they have lags when dealing with continuous environmental changes and are difficult to capture dynamic causal factors in the environment.
[0004] In view of this, an adaptable causal representation decision-making method based on group coding is needed. Summary of the invention
[0005] In view of this, the present invention provides a causal representation decision method based on group coding to improve the adaptability of the causal representation decision method.
[0006] In a first aspect, the present invention provides a causal representation decision method based on population coding, the method comprising: obtaining target environment information; processing the target environment information through a population coding algorithm to obtain a target pulse sequence with time characteristics; processing the target pulse sequence through a target pulse neural network to obtain a target decision probability; based on the target decision probability, outputting a decision result for the target environment information.
[0007] In the embodiment of the present disclosure, the adaptability to dynamic environments can be improved by converting target environment information into a target pulse sequence and inputting the sequence into a pulse neural network to obtain target decision probability and decision results.
[0008] In an optional embodiment, the method also includes: obtaining an initial pulse neural network, wherein the initial pulse neural network includes a plurality of pulse neuron layers; obtaining a target data set, wherein the target data set includes environmental samples and preset output values of each pulse neuron layer; obtaining the actual output value of each pulse neuron layer in the network architecture of the initial pulse neural network when the environmental sample is used as input content; determining the output error of the initial pulse neural network based on the preset output value and the actual output value; adjusting and updating the weight parameters of each pulse neuron layer based on the output error; and determining the target pulse neural network based on the updated weight parameters.
[0009] In the disclosed embodiment, the prediction accuracy of the network can be improved by training the initial spiking neural network with the target data set and adjusting the parameters of the initial spiking neural network.
[0010] In an optional implementation, obtaining the initial spiking neural network includes: selecting corresponding initial network parameters according to target environment information; and generating the initial spiking neural network according to the initial network parameters.
[0011] In the disclosed embodiment, by selecting network parameters according to target environment information and then obtaining an initial spiking neural network, the training time of the initial spiking neural network can be reduced.
[0012] In an optional embodiment, obtaining the initial network parameters includes: constructing a network architecture of a pre-trained network, wherein the number of pulse neuron layers in the pre-trained network corresponds to the number of pulse neuron layers in the target pulse neural network; obtaining pre-training data, wherein the pre-training data is historical environmental information; determining the initial weight parameters of each pulse neuron layer in the network architecture of the pre-trained network; inputting the pre-training data into the pre-training model, and obtaining the first output result of each pulse neuron layer in the pre-training model; based on the first output result, determining the first learning rate of the pre-trained network; based on the first learning rate, adjusting and updating the initial weight parameters of each pulse neuron layer in the pre-trained network; and obtaining the updated initial weight parameters.
[0013] In the embodiments of the present disclosure, by constructing the same network architecture and using pre-training data to train the network to obtain network parameters, the accuracy of the network can be improved.
[0014] In an optional embodiment, the network architecture of the target pulse neural network also includes a pulse gated recurrent unit and a pulse attention layer: the pulse gated recurrent unit is connected after the pulse neuron layer and is used to extract the timing characteristics of the output information of the pulse neuron layer; the pulse attention layer is connected after the pulse neuron layer and is used to extract the attention weight of the output information of the pulse neuron layer.
[0015] In the embodiments of the present disclosure, by adding a pulse gated recurrent unit to the network architecture, it can better adapt to the dynamic environment, and by adding a pulse attention layer to the network architecture, the adaptability and generalization ability of the network can be enhanced.
[0016] In an optional embodiment, the method also includes: constructing a first experience library, the first experience library is used to store historical causal information, wherein the historical causal information includes historical environmental information, historical data encoding, and network parameters of a historical target pulse neural network.
[0017] In the disclosed embodiment, by constructing a first experience library containing historical causal information, the time for model training can be reduced and the generalization ability of the pulse neural network can be improved.
[0018] In an optional implementation, the target pulse sequence with time characteristics includes time delay coding and sequence sparse coding of target environment information.
[0019] In the disclosed embodiment, the target pulse sequence includes time delay coding and sequence sparse coding, which can flexibly and efficiently discover the causal relationship in the environmental information and map these causal relationships into a pulse time sequence.
[0020] In a second aspect, the present invention provides a causal representation decision-making device based on group coding, the device comprising: an environment acquisition module for acquiring target environment information; a sequence acquisition module for processing the target environment information through a group coding algorithm to obtain a target pulse sequence with time characteristics; a probability acquisition module for processing the target pulse sequence through a target pulse neural network to obtain a target decision probability; and an output module for outputting a decision result for the target environment information based on the target decision probability.
[0021] In a third aspect, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the group coding-based causal representation decision method of the above-mentioned first aspect or any corresponding embodiment thereof by executing the computer instructions.
[0022] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the causal representation decision method based on group coding of the above-mentioned first aspect or any corresponding embodiment thereof. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0024] Figure 1 is a flow chart of a causal representation decision method based on group coding according to an embodiment of the present invention;
[0025] Figure 2 is a flow chart of another causal representation decision method based on group coding according to an embodiment of the present invention;
[0026] Figure 3 is a flow chart of another causal representation decision method based on group coding according to an embodiment of the present invention;
[0027] Figure 4 is a structural block diagram of a causal representation decision-making device based on group coding according to an embodiment of the present invention;
[0028] Figure 5 It is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0029] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 those skilled in the art without creative work are within the scope of protection of the present invention.
[0030] SNN (Spiking Neural Network) is the third generation of neural network. It can efficiently process time series information by simulating the pulse emission behavior of biological neurons. Compared with traditional artificial neural networks, SNN has higher computational efficiency and biological interpretability.
[0031] The embodiment of the present invention provides a causal representation decision method based on group coding, which can improve the information processing efficiency and environmental adaptability of the causal representation decision method by combining SNN with causal representation and reasoning cognitive theory.
[0032] According to an embodiment of the present invention, an embodiment of a causal characterization decision-making method based on group coding is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.
[0033] In this embodiment, a causal representation decision method based on group coding is provided. Figure 1 is a flow chart of a causal representation decision method based on group coding according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:
[0034] Step S101, obtaining target environment information.
[0035] The method of obtaining the target environment information is not limited, and corresponding sensors can be set according to the actual situation to collect the environment information. The selection of sensors is not limited, and they can be temperature sensors, humidity sensors, light sensors, sound sensors, etc. Image sensors can also be included to collect videos or images of the surrounding environment.
[0036] Step S102, processing the target environment information through a group coding algorithm to obtain a target pulse sequence with time characteristics.
[0037] The acquisition method of the target pulse sequence with time characteristics is not limited. The acquired target environment information can be preprocessed first, wherein the preprocessing method is not limited and may include data cleaning, data normalization, etc. Then, the target environment information after data preprocessing is feature extracted to obtain the relationship between entities in the surrounding environment, wherein the entity may be a person, a table, etc., and the relationship between entities includes position relationship, interaction relationship, etc. After obtaining the relationship between entities, it is converted into a symbolic form. Different symbols are represented by different neuron groups to represent the target environment information. Then, a pulse emission mechanism is set, wherein the pulse emission mechanism includes a threshold setting and a time window setting for the sensor. The threshold setting is to set a threshold for the sensor that collects the target environment information, and when the collected data exceeds the threshold, a pulse discharge will be triggered. The time window setting is to set a time window, within which the collected data exceeds the threshold to trigger a pulse discharge. Then, the data is mapped to the corresponding time point to obtain a pulse sequence. The mapping method is not limited, and a linear mapping can be selected to directly map the data value to the time point, or a nonlinear function can be used for mapping.
[0038] In a practical application, the target environmental information collected is the cup falling to the ground and the cup breaking. After the "cup falling to the ground" event is detected, the neuron group representing the position of the cup discharges, and a series of pulse sequences representing the event are obtained. The discharge time point in the pulse sequence records the moment of the event, and the density of the pulse sequence records the attributes of the event, such as the material of the cup, the height of the cup falling to the ground, etc. Subsequently, the "cup breaking" event occurs, and after the first time delay, the neuron group representing the shape of the cup discharges, generating another series of pulse sequences. The setting of the first time delay is not restricted, and can be the actual moment of the event, or it can be delayed for a period of time. After the two neuron groups discharge successively, the synaptic connection strength between them will be enhanced, thereby completing the memory of the causal relationship between "the cup falling to the ground" and "the cup breaking".
[0039] Step S103, processing the target pulse sequence through a target pulse neural network to obtain a target decision probability.
[0040] The target pulse neural network includes several layers of pulse neuron layers, and each layer of pulse neuron layers processes the input target pulse sequence.
[0041] In a practical application, the target pulse neural network includes a bottom neuron layer, a middle neuron layer, and an upper neuron layer. After receiving the target pulse sequence of the "cup falling to the ground" and "cup breaking" events, the bottom neuron layer will capture the time point and related attributes of the events, where the related attributes may include the height of the fall, etc. The middle neuron layer will learn the temporal correlation of the target pulse sequence events of the two events and obtain the causal mapping relationship between the two events. The upper neuron will output the probability distribution prediction of "whether the cup will break". Then, based on the probability distribution prediction output by the upper neuron, one or more target decision probabilities can be generated.
[0042] Step S104: outputting a decision result for the target environment information based on the target decision probability.
[0043] Among them, one or more decision thresholds can be set according to actual conditions, and by judging whether the target decision probability reaches the decision threshold, the decision rules are applied to output the decision result for the target environmental information.
[0044] In a practical application, the method is used in a robot dog exploration scene to obtain the optimal decision of the robot dog's obstacle avoidance and walking according to the target decision probability, where the decision includes moving forward, backward, and rotating on the spot. After obtaining the final decision, it can be converted into an actual walking instruction and sent to the robot dog's motion control system to control the actual motion operation of the robot dog.
[0045] The causal representation decision method based on group coding provided in this embodiment can improve adaptability to dynamic environments by converting target environmental information into a target pulse sequence and inputting the sequence into a pulse neural network to obtain target decision probability and decision results.
[0046] In this embodiment, a causal representation decision method based on group coding is provided. Figure 2 is a flow chart of a causal representation decision method based on group coding according to an embodiment of the present invention. Figure 2 As shown, the process includes the following steps:
[0047] Step S201, obtaining target environment information.
[0048] For details, please see Figure 1 Step S101 of the illustrated embodiment will not be described in detail here.
[0049] Step S202, processing the target environment information through a group coding algorithm to obtain a target pulse sequence with time characteristics.
[0050] For details, please see Figure 1 Step S102 of the illustrated embodiment will not be described in detail here.
[0051] Step S203, processing the target pulse sequence through a target pulse neural network to obtain a target decision probability.
[0052] Specifically, the above step S203 includes:
[0053] Step S2031, constructing a target pulse neural network.
[0054] Among them, the construction method of the target pulse neural network is not limited. According to the actual situation, the network structure of the target pulse neural network can be determined first, wherein the network structure includes the number of neuron layers, connection mode, etc. Then select a suitable neuron model and determine the parameters of the neuron, wherein the selection of the neuron model is not limited and can be any neuron model. Then initialize the weight of the synapse, wherein the initialization method is not limited and can be random initialization or preset rules. Then use the training data set to train the target pulse neural network, wherein the training method is not limited and can be supervised learning or unsupervised learning, etc. During the training, adjust the network parameters to optimize the performance of the target pulse neural network. The rules for adjusting the weights are not limited and can be STDP (Spike Timing Dependent Plasticity, synaptic timing dependent plasticity), RSTDP (Reward-Modulated Spike Timing Dependent Plasticity, reverse synaptic timing dependent plasticity), etc.
[0055] In some optional implementations, the above step S2031 includes:
[0056] Step S2031-1, obtaining an initial spiking neural network, wherein the initial spiking neural network includes a plurality of spiking neuron layers.
[0057] There are no restrictions on the number of spiking neuron layers included in the initial spiking neural network, and the connection mode between each spiking neuron layer is also not restricted, and can be fully connected, partially connected, etc. The initialization rule of the synaptic weights in the initial spiking neural network is also not restricted, and can be set randomly, or the initial value of the synaptic weights can be directly set according to actual conditions or historical data.
[0058] In some optional implementations, the above step S2031-1 includes:
[0059] Step a1: Select corresponding initial network parameters according to target environment information.
[0060] The method of obtaining the initial network parameters is not limited, and can be pre-trained network parameters or reference network parameters obtained from relevant websites. Different initial network parameters correspond to different task types and different environments. According to the target environment information, a set of initial network parameters trained on similar causal patterns is selected.
[0061] In a practical application, the initial network parameters are pre-trained using an unsupervised hierarchical sparse coding method. The unsupervised hierarchical sparse coding method draws on the hierarchical structure and coding principles of the visual cortex of the brain, which can improve the learning efficiency of the target spiking neural network.
[0062] In some optional implementations, the above step a1 includes:
[0063] Step a11, constructing a network architecture of a pre-trained network, wherein the number of spiking neuron layers in the pre-trained network corresponds to the number of spiking neuron layers in the target spiking neural network.
[0064] The network architecture of the pre-trained network is exactly the same as that of the target spiking neural network, and the connection method is also the same. In a practical application, the target spiking neural network includes not only a spiking neuron layer, but also an attention layer and a gated recurrent unit, and the network architecture of the pre-trained network also includes an attention layer and a gated recurrent unit.
[0065] Step a12, obtaining pre-training data, wherein the pre-training data is historical environment information;
[0066] Among them, the method of obtaining pre-training data is not restricted and can be directly obtained from relevant websites or platforms.
[0067] Step a13, determining the initial weight parameters of each pulse neuron layer in the network architecture of the pre-trained network.
[0068] There is no restriction on the method of determining the initial weight parameters. The weights may be initialized randomly or other weight initialization methods may be used.
[0069] Step a14, input the pre-training data into the pre-training model, and obtain the first output result of each pulse neuron layer in the pre-training model.
[0070] Step a15: determining a first learning rate of the pre-trained network based on the first output result.
[0071] According to the stability and variation of the first output result, a suitable learning rate is selected as the first learning rate of the pre-trained network.
[0072] In a practical application, if the first output result changes greatly, the first learning rate needs to be appropriately reduced.
[0073] Step a16, based on the first learning rate, adjusting and updating the initial weight parameters of each pulse neuron layer in the pre-trained network.
[0074] There is no restriction on the adjustment method of the initial weight parameters, and the back propagation algorithm and the corresponding optimizer can be used. During the training process, the optimizer can be used to update the weight parameters.
[0075] Step a17, obtaining the updated initial weight parameters.
[0076] After obtaining the updated initial weight parameters, the test set can be used to evaluate and train the performance of the network model. Based on the test results, the initial weight parameters can also be adjusted and updated for a second time.
[0077] By building the same network architecture and using the network parameters obtained by training the network with pre-training data, the accuracy of the network can be improved.
[0078] Step a2, generating an initial pulse neural network according to the initial network parameters.
[0079] By selecting network parameters according to target environment information and then obtaining an initial spiking neural network, the training time of the initial spiking neural network can be reduced.
[0080] Step S2031 - 2 , obtaining a target data set, wherein the target data set includes environmental samples and preset output values of each pulse neuron layer.
[0081] There is no restriction on the method of obtaining the target data set, which can be directly obtained through a related platform or website, or environmental samples can be collected or generated from the target environment, and then the preset output value of the pulse neuron layer is determined for each environmental sample, wherein the method of determining the preset output value is not restricted, and can be obtained based on prior knowledge or manual annotation. After obtaining the target data set, the target data set can be divided into a training set, a validation set, and a test set.
[0082] Step S2031-3, obtaining the actual output value of each pulse neuron layer in the network architecture of the initial pulse neural network when the environmental sample is used as input content.
[0083] There is no restriction on the way of obtaining the actual output value, and the input data can be forward propagated through the initial pulse neural network to obtain the actual output value of the network.
[0084] Step S2031-4, determining the output error of the initial pulse neural network based on the preset output value and the actual output value.
[0085] Among them, the method of obtaining the output error is not restricted and can be obtained by selecting a suitable loss function.
[0086] Step S2031-5, based on the output error, adjust and update the weight parameters of each pulse neuron layer.
[0087] The gradient of the loss function can be obtained according to the output error, and then the weight parameters can be updated according to the gradient to reduce the error. The weight parameters of each spiking neuron layer can be adjusted by the optimizer.
[0088] Step S2031-6, determining the target pulse neural network based on the updated weight parameters.
[0089] Step S2032, input the target pulse sequence into the target pulse neural network to obtain the target decision probability.
[0090] Step S204: outputting a decision result for the target environment information based on the target decision probability.
[0091] For details, please see Figure 1 Step S104 of the illustrated embodiment will not be described in detail here.
[0092] The causal representation decision method based on population coding provided in this embodiment trains the initial spiking neural network through the target data set and adjusts the parameters of the initial spiking neural network, thereby improving the prediction accuracy of the network.
[0093] In this embodiment, a causal representation decision method based on group coding is provided. Figure 3 is a flow chart of a causal representation decision method based on group coding according to an embodiment of the present invention. Figure 3 As shown, the process includes the following steps:
[0094] Step S301, obtaining target environment information.
[0095] For details, please see Figure 1 Step S101 of the illustrated embodiment will not be described in detail here.
[0096] Step S302: Process the target environment information through a group coding algorithm to obtain a target pulse sequence with time characteristics.
[0097] In some optional embodiments, the target pulse sequence with time characteristics includes time delay coding and sequence sparse coding of target environment information.
[0098] The target pulse sequence includes time delay coding and sequence sparse coding, which can flexibly and efficiently discover causal relationships in environmental information and map these causal relationships into pulse time sequences.
[0099] Step S303, processing the target pulse sequence through a target pulse neural network to obtain a target decision probability.
[0100] For details, please see Figure 1 Step S103 of the illustrated embodiment will not be described in detail here.
[0101] In some optional embodiments, the network architecture of the target pulse neural network also includes a pulse gated recurrent unit and a pulse attention layer: the pulse gated recurrent unit is connected after the pulse neuron layer, and is used to extract the timing characteristics of the output information of the pulse neuron layer; the pulse attention layer is connected after the pulse neuron layer, and is used to extract the attention weight of the output information of the pulse neuron layer.
[0102] The location of the pulse gated recurrent unit is not restricted, and it can be used as the last layer of the target pulse neural network or connected to the pulse neuron layer. The method for determining the initial value of the pulse gated recurrent unit is not restricted, and it can be randomly initialized or preset according to actual conditions. The method for adjusting the synaptic weight of the pulse gated recurrent unit is not restricted, and a suitable learning rule can be selected according to actual conditions.
[0103] The maintenance and number of pulse attention layers are not limited. They can be set after each pulse neuron layer to obtain the attention weight of each pulse neuron layer. The setting of the attention mechanism is not limited. The importance of different pulses can be obtained according to the attention weight of each pulse neuron layer. The initial weight setting of the pulse attention layer is not limited. The appropriate initial value can be selected according to the actual situation. Then training is performed to optimize the weight of the pulse attention layer, thereby improving the efficiency and accuracy of network information processing.
[0104] By adding pulse-gated recurrent units to the network architecture, it can better adapt to dynamic environments. By adding pulse attention layers to the network architecture, the adaptability and generalization ability of the network can be enhanced.
[0105] Step S304: outputting a decision result for the target environment information based on the target decision probability.
[0106] For details, please see Figure 1 Step S104 of the illustrated embodiment will not be described in detail here.
[0107] Step S305, constructing a first experience database, the first experience database is used to store historical causal information, wherein the historical causal information includes historical environmental information, historical data encoding, and network parameters of the historical target pulse neural network.
[0108] There is no restriction on the way of constructing the first experience database, and each time the target environment information is collected and the target data code and the network parameters of the target pulse neural network corresponding to the target environment information are obtained, they can be saved in the first experience database. There is no restriction on the representation of the historical causal information in the first experience database, and after obtaining the target environment information, the target data code and the network parameters, they can be integrated into a unified data structure.
[0109] In a practical application, the scenario is intelligent patrol. When the robot dog enters a new terrain environment, it obtains new target environment information, where the target environment information may include "obstacles detected", "flat ground" and "slope", etc. When constructing the target spiking neural network, the network parameters most similar to these environmental information can be found from the first experience library to initialize the target spiking neural network.
[0110] The causal representation decision-making method based on group coding provided in this embodiment can reduce the time of model training and improve the generalization ability of the pulse neural network by constructing a first experience library containing historical causal information.
[0111] In this embodiment, a causal characterization decision-making device based on group coding is also provided, which is used to implement the above-mentioned embodiments and preferred implementation modes, and will not be repeated hereafter. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.
[0112] This embodiment provides a causal representation decision-making device based on group coding, such as Figure 4 As shown, including:
[0113] The environment acquisition module 401 is used to acquire target environment information.
[0114] The sequence acquisition module 402 is used to process the target environment information through a group coding algorithm to obtain a target pulse sequence with time characteristics.
[0115] The probability acquisition module 403 is used to process the target pulse sequence through the target pulse neural network to obtain the target decision probability.
[0116] The output module 404 is used to output a decision result for the target environment information based on the target decision probability.
[0117] In some optional implementations, the causal representation decision-making device based on group coding further includes:
[0118] The target spiking neural network construction module 405 is used to construct the target spiking neural network.
[0119] The target spiking neural network building module 405 includes:
[0120] The initial network acquisition unit is used to acquire an initial spiking neural network, wherein the initial spiking neural network includes a plurality of spiking neuron layers.
[0121] The target data set acquisition unit is used to acquire the target data set, wherein the target data set includes the environment sample and the preset output values of each pulse neuron layer.
[0122] The output acquisition unit is used to obtain the actual output value of each pulse neuron layer in the network architecture of the initial pulse neural network when the environmental sample is used as input content.
[0123] The error determination unit is used to determine the output error of the initial pulse neural network based on the preset output value and the actual output value.
[0124] The weight adjustment unit is used to adjust and update the weight parameters of each spiking neuron layer based on the output error.
[0125] The target spiking neural network determining unit is used to determine the target spiking neural network based on the updated weight parameters.
[0126] In some optional implementations, the initial network acquisition unit includes:
[0127] The initial network parameter acquisition subunit is used to select corresponding initial network parameters according to the target environment information.
[0128] The initial pulse neural network generation subunit is used to generate an initial pulse neural network according to the initial network parameters.
[0129] In some optional implementations, the initial network parameter acquisition subunit includes:
[0130] A pre-trained network architecture building component is used to build a network architecture of a pre-trained network, where the number of spiking neuron layers in the pre-trained network corresponds to the number of spiking neuron layers in the target spiking neural network.
[0131] The pre-training data acquisition component is used to acquire pre-training data, where the pre-training data is historical environment information.
[0132] The initial weight parameter determination component is used to determine the initial weight parameters of each pulse neuron layer in the network architecture of the pre-trained network.
[0133] The output result acquisition component is used to input the pre-training data into the pre-training model and obtain the first output result of each pulse neuron layer in the pre-training model.
[0134] The learning rate determination component is used to determine a first learning rate of the pre-trained network based on the first output result.
[0135] The weight updating component is used to adjust and update the initial weight parameters of each pulse neuron layer in the pre-trained network based on the first learning rate.
[0136] The update parameter acquisition component is used to obtain the updated initial weight parameters.
[0137] In some optional implementations, the target spiking neural network building module 405 further includes:
[0138] The experience library construction unit is used to construct a first experience library, and the first experience library is used to store historical causal information, wherein the historical causal information includes historical environmental information, historical data encoding, and network parameters of the historical target pulse neural network.
[0139] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.
[0140] The group coding-based causal representation decision-making device in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.
[0141] The embodiment of the present invention also provides a computer device having the above Figure 4 The group coding-based causal representation decision-making device shown.
[0142] See also Figure 5 , Figure 5 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Figure 5As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 5 A processor 10 is taken as an example.
[0143] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.
[0144] The memory 20 stores instructions executable by at least one processor 10, so that at least one processor 10 executes the method shown in the above embodiment.
[0145] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.
[0146] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.
[0147] The computer device also includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Figure 5 The example of connecting through bus is taken in the following.
[0148] The input device 30 can receive input digital or character information, and generate key signal input related to the user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, an indicator bar, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (e.g., an LED) and a tactile feedback device (e.g., a vibration motor), etc. The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display and a plasma display. In some optional embodiments, the display device can be a touch screen.
[0149] The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.
[0150] A part of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the existence of the computer program instruction in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc., and accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium accessible to the computer.
[0151] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.
Claims
1. A causal representation decision method based on group coding, It is characterized in that The method comprises: Obtain target environment information; Processing the target environment information through a group coding algorithm to obtain a target pulse sequence with time characteristics; Processing the target pulse sequence through a target pulse neural network to obtain a target decision probability; Based on the target decision probability, a decision result for the target environment information is output.
2. The method according to claim 1, It is characterized in that The method further comprises: Acquire an initial spiking neural network, wherein the initial spiking neural network includes a plurality of spiking neuron layers; Acquire a target data set, wherein the target data set includes an environmental sample and preset output values of each spiking neuron layer; Acquire the actual output value of each spiking neuron layer in the network architecture of the initial spiking neural network when the environmental sample is used as input content; Determining an output error of the initial spiking neural network based on the preset output value and the actual output value; Based on the output error, adjusting and updating the weight parameters of each spiking neuron layer; Based on the updated weight parameters, the target spiking neural network is determined.
3. The method according to claim 2, It is characterized in that The obtaining of the initial spiking neural network comprises: According to the target environment information, select corresponding initial network parameters; The initial spiking neural network is generated according to the initial network parameters.
4. The method according to claim 3, It is characterized in that The acquisition of the initial network parameters includes: Constructing a network architecture of a pre-trained network, wherein the number of spiking neuron layers in the pre-trained network corresponds to the number of spiking neuron layers in the target spiking neural network; Acquire pre-training data, wherein the pre-training data is historical environment information; Determining initial weight parameters of each spiking neuron layer in the network architecture of the pre-trained network; Inputting the pre-training data into the pre-training model, and obtaining the first output result of each spiking neuron layer in the pre-training model; Determining a first learning rate of the pre-trained network based on the first output result; Based on the first learning rate, adjusting and updating the initial weight parameters of each spiking neuron layer in the pre-trained network; Get the updated initial weight parameters.
5. The method according to claim 2, It is characterized in that The network architecture of the target spiking neural network also includes a spiking gated recurrent unit and a spiking attention layer: The pulse gated recurrent unit is connected after the pulse neuron layer and is used to extract the timing characteristics of the output information of the pulse neuron layer; The pulse attention layer is connected after the pulse neuron layer and is used to extract the attention weight of the output information of the pulse neuron layer.
6. The method according to claim 2, It is characterized in that The method further comprises: Construct a first experience library, which is used to store historical causal information, wherein the historical causal information includes historical environmental information, historical data encoding, and network parameters of a historical target pulse neural network.
7. The method according to claim 1, It is characterized in that The target pulse sequence with time characteristics includes time delay coding and sequence sparse coding of the target environment information.
8. A causal representation decision-making device based on group coding, It is characterized in that The device comprises: Environment acquisition module, used to obtain target environment information; A sequence acquisition module, used to process the target environment information through a group coding algorithm to obtain a target pulse sequence with time characteristics; A probability acquisition module, used for processing the target pulse sequence through a target pulse neural network to obtain a target decision probability; An output module is used to output a decision result for the target environment information based on the target decision probability.
9. A computer device, It is characterized in that include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the causal representation decision method based on group coding according to any one of claims 1 to 7 by executing the computer instructions.
10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the causal representation decision method based on population coding according to any one of claims 1 to 7.
