Brain-like synapse learning method and brain-like technology neuromorphic hardware system

Abstract

The application provides a brain-like synapse learning method and a brain-like technology neural morphological hardware system, and the method comprises the following steps: determining a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in a brain-like synapse learning circuit, wherein the pulse pair comprises an input pulse generated by the presynaptic neuron and an output pulse generated by the postsynaptic neuron; determining an STDP mechanism corresponding to the pulse pair and a synapse weight corresponding to the STDP mechanism based on the pulse pair; and performing STDP learning corresponding to the brain-like synapse learning circuit based on the pulse pair and the synapse weight; wherein the STDP mechanism is a pulse time-dependent plasticity mechanism, and the front and rear pulses of the pulse pair correspond to a long-term potentiation process or a long-term depression process in the STDP mechanism according to the time sequence. The application realizes online learning of brain-like intelligence, and plays the environment self-adaptive characteristics of brain-like computing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a brain-like synapse learning method and a neuromorphic hardware system of brain-like technology. BACKGROUND

[0002] The powerful perception, learning and adaptation ability of the human brain is an important reference object for people in the process of realizing strong artificial intelligence. A certain on-chip learning ability is indispensable for a neuromorphic hardware platform to serve as a truly intelligent system. At present, there is no universally recognized best algorithm for online learning algorithm selection. On the one hand, we still do not have a good understanding of the learning mechanism of the brain, and on the other hand, the application scenarios of neuromorphic brain platforms are also very wide, and various learning algorithms more or less have certain limitations, and there is no all-around algorithm suitable for all tasks. Therefore, how to establish a better online learning mechanism has become a heavy and difficult point at present.

[0003] In the existing technology, the most direct "take it as it is" ANN to SNN scheme is used to realize offline learning. In this way, the chip actually does not have learning ability, and it is a direct deployment of the achievements of traditional ANN to neuromorphic hardware. The limitation it faces is that the learning effect is completely limited by the pre-ANN scheme, and it does not have the further learning ability of the human brain. When switching to the SNN model, the implementation effect is still lossy, which does not meet the current development trend of online learning.

[0004] Some technologies are committed to realizing configurable online learning algorithms, such as the Loihi chip equipped with a programmable learning engine, which can realize complex and diverse learning mechanisms. The disadvantage is that the cost of implementing learning is high, and the pursuit of universality leads to high learning energy consumption and limited computing power.

[0005] And the pulse frequency coding method in some other algorithms like SDSP rule and Spike Count rule cannot fully represent the actual role of synaptic plasticity. The reinforcement learning circuit of the e-prop rule requires a relatively complex reward and punishment signal, which increases the design difficulty, and the learning efficiency is too low.

[0006] Therefore, due to the defects of high learning energy consumption, limited computing power, and limited application scenarios of existing chips or circuits executing online learning algorithms, establishing a better online learning mechanism has become a heavy and difficult point at present. SUMMARY

[0007] The application provides a brain-like synapse learning method and a neuromorphic hardware system of brain-like technology, which aims to solve the defects of the existing neuromorphic hardware system or chip or circuit of brain-like technology in high learning energy consumption, limited computing power and limited application scenarios, thereby improving the online learning of brain-like intelligence and playing the environment adaptive characteristics of brain-like computing.

[0008] The application provides a brain-like synapse learning method, which is applied to a neuromorphic hardware system of brain-like technology, and the neuromorphic hardware system comprises a brain-like synapse learning circuit, and the method comprises the following steps:

[0009] determining a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in the brain-like synapse learning circuit, wherein the pulse pair comprises an input pulse generated by the presynaptic neuron and an output pulse generated by the postsynaptic neuron;

[0010] based on the pulse pair, determining an STDP mechanism corresponding to the pulse pair and a synaptic weight corresponding to the STDP mechanism;

[0011] based on the pulse pair and the synaptic weight, performing STDP learning corresponding to the brain-like synapse learning circuit;

[0012] wherein the STDP mechanism is a pulse time-dependent plasticity mechanism, the brain-like synapse learning circuit comprises a synapse, the presynaptic neuron and the postsynaptic neuron, and the front and rear pulses of the pulse pair correspond to a long-term potentiation process or a long-term depression process in the STDP mechanism according to the time sequence.

[0013] According to the brain-like synapse learning method provided by the application, based on the pulse pair, the synaptic weight corresponding to the STDP mechanism is determined, which comprises the following steps:

[0014] based on the pulse pair, determining the nearest neighbor peak value of the pulse pair;

[0015] based on the nearest neighbor peak value and the weight amplitude generated by the STDP mechanism, determining the synaptic weight corresponding to the STDP mechanism.

[0016] According to the brain-like synapse learning method provided by the application, the weight amplitude generated by the STDP mechanism is determined based on the following steps:

[0017] based on the time point at which the pulse pair is generated, determining the time difference between the front and rear pulses of the pulse pair;

[0018] if the time difference is greater than or equal to a preset threshold, it is determined that the weight amplitude is generated by the long-term potentiation process;

[0019] If the time difference is less than the preset threshold, the weight amplitude is determined as generated by the long-term depression process.

[0020] According to the brain-like synaptic learning method provided by the application, the input pulse is determined by the following steps:

[0021] An input pulse sample is obtained.

[0022] The input pulse sample is weighted by a Poisson distribution process to obtain a preprocessed sample corresponding to the input pulse sample.

[0023] Based on a preset time window, the number and proportion of pulses of the preprocessed sample within the preset time window are determined.

[0024] The Poisson distribution process weighting is used to weight the input pulse sample based on a weighting value of the Poisson distribution, and the preprocessed sample is used to represent the input pulse in terms of the number of pulses and the proportion.

[0025] According to the brain-like synaptic learning method provided by the application, the input pulse is determined by the following steps:

[0026] An input pulse sample is obtained.

[0027] The input pulse sample is time-encoded so that each pixel value of the input pulse sample corresponds to a pulse, and the size of each pixel value of the input pulse sample is mapped to a specific time in a preset time window to obtain the input pulse.

[0028] According to the brain-like synaptic learning method provided by the application, the pulse pair is determined by the following steps:

[0029] If it is detected that any neuron in the brain-like synaptic learning circuit emits a pulse, a global inhibition signal is generated, and the global inhibition signal is used to inhibit the pulse firing activity of all other neurons.

[0030] Based on the inhibition amplitude corresponding to the global inhibition signal, the pulse pair is determined.

[0031] The application also provides a neuro-morphic hardware system of brain-like technology, comprising:

[0032] A pulse input module is used to determine a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in a brain-like synaptic learning circuit, the pulse pair comprising an input pulse generated by the presynaptic neuron and an output pulse generated by the postsynaptic neuron.

[0033] a synaptic weight calculation module, configured to determine an STDP mechanism corresponding to the pulse pair and a synaptic weight corresponding to the STDP mechanism based on the pulse pair;

[0034] a brain-like synaptic learning module, configured to perform STDP learning of the brain-like synaptic learning circuit based on the pulse pair and the synaptic weight;

[0035] The STDP mechanism is a pulse time-dependent plasticity mechanism, the brain-like synaptic learning circuit comprises a synapse, the presynaptic neuron and the postsynaptic neuron, and the pre-pulse and the post-pulse of the pulse pair correspond to a long-term potentiation process or a long-term depression process in the STDP mechanism according to a time sequence.

[0036] The application further provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the brain-like synaptic learning method according to any of the above when executing the program.

[0037] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the brain-like synaptic learning method according to any of the above.

[0038] The application further provides a computer program product comprising a computer program, wherein the computer program is executable on a processor to implement the brain-like synaptic learning method according to any of the above.

[0039] The brain-like synaptic learning method and the neuromorphic hardware system of brain-like technology provided by the application perform STDP learning of a brain-like synaptic learning circuit according to a pulse pair generated by a synaptic neuron and a synaptic weight corresponding to the pulse pair, and the pre-pulse and the post-pulse correspond to a long-term potentiation process or a long-term depression process in the STDP mechanism. Since the complexity of the STDP learning rule is moderate, the update rule can reflect the long-term potentiation process or the long-term depression process in the synaptic learning process of a living being, and the brain-like synaptic learning circuit has the advantages of being bionic, lightweight and locally parallel. The application proposes a brain-like synaptic learning circuit with pulse time-dependent plasticity, which can achieve an STDP learning engine with the best learning energy efficiency by weighing the reasonable interpretability of a living being and the efficient online learning effect, so as to equip a chip with actual online learning ability and make the brain-like computing platform have the environment adaptive characteristics. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to make the technical solutions in the present application or prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative work based on the accompanying drawings also belong to the protection scope of the present application.

[0041] Figure 1 is one of the flowcharts of the brain-like synapse learning method provided by the present application;

[0042] Figure 2 is another flowchart of the brain-like synapse learning method provided by the present application;

[0043] Figure 3 is a third flowchart of the brain-like synapse learning method provided by the present application;

[0044] Figure 4 is a fourth flowchart of the brain-like synapse learning method provided by the present application;

[0045] Figure 5 is a fifth flowchart of the brain-like synapse learning method provided by the present application;

[0046] Figure 6 is a sixth flowchart of the brain-like synapse learning method provided by the present application;

[0047] Figure 7 is a structural schematic diagram of the neuromorphic hardware system of the brain-like technology provided by the present application;

[0048] Figure 8 is a structural schematic diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0049] In order to make the technical solutions in the present application or prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative work based on the accompanying drawings also belong to the protection scope of the present application.

[0050] The brain-like synapse learning method of the present application will be described below. Figures 1-6

[0051] The brain-like synapse learning method provided by the present application comprises: Figure 1

[0052] ​​Step 10, determining a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in the brain-like synapse learning circuit, the pulse pair including an input pulse generated by the presynaptic neuron and an output pulse generated by the postsynaptic neuron.

[0053] The pulse pair carries the neural feedback information generated by the brain-like synapse learning circuit for the input signal, including the input pulse generated by the presynaptic neuron and the output pulse generated by the postsynaptic neuron.

[0054] It should be noted that the brain-like synapse learning method proposed by the application is applied to a neuromorphic hardware system of brain-like technology, and the neuromorphic hardware system includes a brain-like synapse learning circuit. The data processing process can be image learning or classification, or voice data processing, or language learning or recognition translation, etc.

[0055] The brain-like synapse learning circuit is a large-scale parallel system composed of multiple neurons and synaptic connections therebetween. The upper-level neuron can transmit signals to the lower-level neuron at the synaptic connection point via the axon. The brain-like synapse learning circuit which is the minimum unit of complete pulse response can include a presynaptic neuron, a synapse, and a postsynaptic neuron. The effect of the pulse sent by the presynaptic neuron to the receiving neuron depends on the strength of the synapse connecting the two neurons, and the synaptic strength and connection mode between neurons have an important influence on the information processing capacity of the nervous system.

[0056] In one embodiment, an image signal is received, and based on the image signal, a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in the brain-like synapse learning circuit is determined. The brain-like synapse learning circuit performs image processing and analysis according to the input image signal, and the presynaptic neuron and the postsynaptic neuron in the brain-like synapse learning circuit generate neural feedback according to the image signal to obtain a pulse pair generated by the neurons in the brain-like synapse learning circuit, wherein the pulse pair carries the neural feedback information generated by the brain-like synapse learning circuit for the image signal.

[0057] Step 20, based on the pulse pair, determining an STDP mechanism corresponding to the pulse pair, and a synaptic weight corresponding to the STDP mechanism;

[0058] Based on the neural feedback information, i.e. the pulse pair, generated by the brain-like synapse learning circuit, the STDP mechanism corresponding to the pulse pair is first determined, and then the synaptic weight corresponding to the STDP mechanism is calculated based on the pulse pair and the STDP mechanism. The synaptic weight represents the connection strength between the input pulse and the output pulse in the pulse pair, and is used to represent the importance degree of the feedback of the pulse pair perceived by the presynaptic neuron and the postsynaptic neuron.

[0059] Step 30, based on the pulse pair and the synaptic weight, the corresponding STDP learning of the brain-like synaptic learning circuit is performed.

[0060] Wherein, the STDP mechanism is a pulse time-dependent plasticity mechanism, the brain-like synaptic learning circuit comprises a synapse, the presynaptic neuron and the postsynaptic neuron, and the pre-pulse and the post-pulse of the pulse pair correspond to the long-term potentiation process or the long-term depression process in the STDP mechanism according to the time sequence.

[0061] The pre-pulse and the post-pulse of the pulse pair can be divided into the long-term potentiation process or the long-term depression process corresponding to the STDP mechanism according to the time sequence. Specifically, after the pulse emitted by the presynaptic neuron reaches the synapse, the postsynaptic neuron generates a response and emits a pulse, then the synaptic weight between the neurons is enhanced, which is called the long-term potentiation process, and correspondingly, if the emission time of the input pulse corresponding to the presynaptic neuron is earlier than the emission time of the output pulse corresponding to the postsynaptic neuron, the pre-pulse and the post-pulse of the pulse pair correspond to the long-term potentiation process in the STDP mechanism. If the postsynaptic neuron emits a pulse after the presynaptic neuron transmits a pulse, the information may be ignored, and then the synaptic weight between the neurons is weakened, which is called the long-term depression process, and correspondingly, if the emission time of the input pulse corresponding to the presynaptic neuron is later than the emission time of the output pulse corresponding to the postsynaptic neuron, the pre-pulse and the post-pulse of the pulse pair correspond to the long-term depression process in the STDP mechanism.

[0062] The pulse pair is distinguished according to the STDP mechanism, and the corresponding STDP learning of the brain-like synaptic learning circuit is performed according to the pulse pair distinguished according to the STDP mechanism and the synaptic weight corresponding to the pulse pair, so as to realize the STDP learning of the brain-like synaptic learning circuit.

[0063] Some technologies are dedicated to realizing configurable online learning algorithms, such as the Loihi chip which is equipped with a programmable learning engine and can realize complex and diverse learning mechanisms, and the disadvantage is that the cost of realizing learning is high, and the pursuit of universality leads to high learning energy consumption and limited computing power.

[0064] And the pulse frequency encoding mode in some algorithms such as the SDSP rule and the Spike Count rule cannot fully represent the actual role of synaptic plasticity, and the reinforcement learning circuit in the e-prop rule needs a relatively complex reward and punishment signal, which increases the design difficulty. In comparison, the complexity of the STDP learning rule is moderate, and the update rule is sufficient to reflect the biological synaptic learning process. It has the advantages of bionics, lightweight, local parallel execution, which promotes the application of the STDP learning algorithm to the basic principle of online learning in the present application.

[0065] To this end, the application provides a brain-like synapse learning circuit with pulse time-dependent plasticity, which is designed based on the trade-off between the biological reasonable explanatory basis and the efficient online learning effect, and can realize the STDP learning engine with the best learning energy efficiency, thereby providing the chip with actual online learning capability and making the brain-like computing platform have the environment adaptive characteristics.

[0066] The brain-like synapse learning method provided by the application performs the STDP learning of the brain-like synapse learning circuit according to the pulse pair generated by the synapse neuron and the synaptic weight corresponding to the pulse pair, and the pre-pulse and the post-pulse correspond to the long-term potentiation process or the long-term depression process in the STDP mechanism. Since the complexity of the STDP learning rule is moderate, the update rule can reflect the long-term potentiation process or the long-term depression process in the biological synapse learning process, and the brain-like synapse learning circuit has the advantages of being bionic, lightweight, and locally parallel. The application provides a brain-like synapse learning circuit with pulse time-dependent plasticity, which is designed based on the trade-off between the biological reasonable explanatory basis and the efficient online learning effect, and can realize the STDP learning engine with the best learning energy efficiency, thereby providing the chip with actual online learning capability and making the brain-like computing platform have the environment adaptive characteristics.

[0067] In an embodiment, referring to Figure 2 , step 20, determining the synaptic weight corresponding to the STDP mechanism based on the pulse pair comprises:

[0068] Step 21, determining the nearest-neighbor peak value of the pulse pair based on the pulse pair.

[0069] Step 22, determining the synaptic weight corresponding to the STDP mechanism based on the nearest-neighbor peak value and the weight amplitude generated by the STDP mechanism.

[0070] The STDP mechanism corresponding to the pulse pair comprises the long-term potentiation process or the long-term depression process.

[0071] The pulse time-dependent plasticity STDP mechanism used in the application was first discovered in the animal brain, and has a clear biological theoretical basis. This biological process is usually described as two cases of long-term potentiation (LTP) and long-term depression (LTD). When the pulses emitted by the pre-neuron and the post-neuron have a certain time sequence, the synaptic weight will change accordingly. When the pre-neuron emits earlier than the post-neuron, it can be considered that the two have a certain degree of correlation, and the pre-neuron causes the post-neuron to emit, so the connection relationship will be enhanced, and the change amount ΔW of the synaptic weight is positive; on the contrary, when the post-neuron emits earlier than the pre-neuron, it is considered that there is no correlation, and the connection relationship will be weakened, and ΔW is negative. In terms of correlation, adjacent pulses can be considered to have stronger correlation, so both the LTP and the LTD curves show exponential decay characteristics.

[0072] In this STDP learning rule based on pulse pairs, if the updated weight depends on the cumulative effect of all past spikes, it is defined as an All-to-all rule; on the contrary, if only the most recent spike is calculated, it is called a nearest neighbor rule. In addition, if the updated weight depends on the current weight value for multiplication, too small or too large weight values will only cause a small change, and it is easy to form a unimodal distribution, that is, Multi-STDP; on the other hand, if the weight can be freely changed within the upper and lower boundaries, it will usually form a bimodal distribution, that is, Add-STDP. All-to-all and Multi-STDP are introduced to fully exploit the input data and maintain the steady state between neurons. However, this requires the integration of a huge spike history buffer and a high-cost multiplier. Considering that the purpose of chip design is to promote low-cost neuromorphic intelligence, the nearest neighbor rule is used to store the minimum pulse history in the design, and Add-STDP is used to avoid any multiplication and division.

[0073] Therefore, the additive nearest neighbor STDP update rule used in the application can be expressed as follows:

[0074]

[0075] where w original is the original synaptic weight, w upper and w lower are the upper and lower hard boundaries of the weight, Δw LTP is the weight amplitude corresponding to the long-term potentiation process, Δw LTD is the weight amplitude corresponding to the long-term depression process, and Δt is the time difference between the post-synaptic neuron output pulse and the pre-synaptic neuron input pulse.

[0076] The embodiment proposes to add the peak value of the nearest neighbor and the weight amplitude generated by the STDP mechanism according to the additive nearest neighbor STDP update rule, to calculate the synaptic weight corresponding to the STDP mechanism, without integrating a huge calculation cumbersome pulse history cache and a high-cost multiplier, to realize a low-cost neuromorphic intelligent circuit, so that the chip cost is greatly reduced.

[0077] In one embodiment, referring to Figure 3 , step 22, determining the weight amplitude generated by the STDP mechanism comprises:

[0078] Step 221, determining the time difference before the pre-pulse and the post-pulse of the pulse pair based on the time when the pulse pair is generated;

[0079] Step 222, if the time difference is greater than or equal to a preset threshold, determining that the weight amplitude is generated by the long-term potentiation process;

[0080] Step 223, if the time difference is less than the preset threshold, determining that the weight amplitude is generated by the long-term depression process.

[0081] Wherein, the STDP mechanism comprises a long-term potentiation process or a long-term depression process, and the configurable lookup table is used to determine that the weight amplitude is generated by the long-term potentiation process if the time difference is greater than or equal to a preset threshold, and to determine that the weight amplitude is generated by the long-term depression process if the time difference between the pre-pulse and the post-pulse is less than the preset threshold.

[0082] To realize the STDP learning weight update, the present application adopts a design scheme combining presynaptic and postsynaptic counters with a lookup table (LUT) in specific circuit implementation. Both presynaptic and postsynaptic neurons will generate corresponding input pulses and output pulses at specific time steps, and recording the accurate time when the pulse pair is generated is the key to realizing STDP learning.

[0083] A 5-bit simple counter can be used to realize the function of recording the pulse. First, all the pulse counters before and after the synapse are reset to the maximum value, and the presynaptic pulse counter is cleared when the presynaptic pulse arrives. In the next time step, if there is no pulse, the counter will be incremented one by one until the maximum value, that is, the counter actually records the time length from the arrival of the last pulse to the current time step. Similarly, the postsynaptic pulse counter is also processed according to whether the neuron generates a pulse.

[0084] When the pre-pulse and post-pulse are read from the presynaptic and postsynaptic pulse counters, their time difference Δt = Counter qre -Counterpost where Counter pre is the time point when the pre-synaptic neuron generates an input pulse, Counter post is the time point when the post-synaptic neuron generates an output pulse. The next question is how to calculate the corresponding weight update amplitude Δw LTP and Δw LTD . In the design, a configurable look-up table (LUT) with 60 words of 8 bits is implemented. In the classical case, the first 30 words are used to index the long-term depression (LTD) process, and the last 30 words are used to index the long-term potentiation (LTP) process. The address of the index is exactly Δt.

[0085] In order to avoid repeated updates, the circuit only performs STDP learning at the time of the arrival of the pre-pulse and the post-pulse, which can distinguish four cases: only the pre-synaptic pulse, only the post-synaptic pulse, both the pre-synaptic pulse and the post-synaptic pulse, and neither the pre-synaptic pulse nor the post-synaptic pulse. When only the pre-synaptic pulse exists, long-term depression (LTD) is performed. Obviously, the pre-synaptic counter is 0, and the post-synaptic counter is not 0. Therefore, -Δt can be used for the long-term depression (LTD) index. When only the post-synaptic pulse exists, long-term potentiation (LTP) is performed. Obviously, the post-synaptic counter is 0, and the pre-synaptic counter is not 0. Therefore, Δt can be used for the long-term potentiation (LTP) index. For the case of both the pre-synaptic pulse and the post-synaptic pulse, long-term potentiation (LTP) is performed because the post-synaptic pulse is the result of the accumulation of the pre-synaptic pulse and has correlation. If neither the pre-synaptic pulse nor the post-synaptic pulse exists, the LUT table is not searched, and STDP learning is not performed. Referring to the aforementioned additive STDP scheme, the original weight w original is added to the weight update amplitude Δw LTP or Δw LTD obtained by the table lookup to obtain the updated weight w updated .

[0086] In addition, considering that the learning curve actually plays a role similar to the learning rate in STDP learning, in order to support a smaller learning rate, the chip also uses a linear feedback shift register to provide random learning capability, that is, the update amplitude of learning is ±1, but whether to update needs to be determined by comparing the index value of the LUT with the pseudo-random number. In terms of effect, the design can achieve fractional probability random STDP update.

[0087] In the design also using the STDP learning algorithm, the embodiment also uses a low-cost implementation scheme of combining the pre-synaptic and post-synaptic counters with the look-up table (LUT). In the same comparison, the design can achieve the best energy efficiency, which is the result of weighing the reasonable explanatory basis of biology and the efficient online learning effect.

[0088] In the present application, a scheme is proposed to implement STDP learning using both emissivity coding and space-time coding.

[0089] In an embodiment, referring to Figure 4 , step 10, the input pulse is determined by the following steps:

[0090] Step 101, obtaining an input pulse sample;

[0091] Step 102, performing Poisson distribution process weighting on the input pulse sample to obtain a preprocessed sample corresponding to the input pulse sample;

[0092] Step 103, determining the number and proportion of pulses of the preprocessed sample within a preset time window based on the preset time window;

[0093] The Poisson distribution process weighting is used to weight the input pulse sample based on a weighting value of Poisson distribution, and the number and proportion of pulses of the preprocessed sample within the preset time window are used to represent the input pulse.

[0094] In the emissivity coding scheme, the input sample is represented by the number and proportion of pulses within a certain time window, and in this case, the learning effect of STDP is similar to that of SDSP learning. In the position with high pixel value, the frequency of pulse emission is obviously higher, and for different pixel value sums of the input sample, there is a weighted Poisson distribution process, so that all samples can be learned more fairly.

[0095] In an embodiment, referring to Figure 5 , step 10, the input pulse is determined by the following steps:

[0096] Step 111, obtaining an input pulse sample;

[0097] Step 112, performing time coding on the input pulse sample, so that each pixel value of the input pulse sample corresponds to a pulse, and the size of each pixel value of the input pulse sample is mapped to a specific time in a preset time window to obtain the input pulse.

[0098] In the time coding scheme, each pixel value corresponds to only one pulse, and the pixel value size is mapped to a specific time in the time window. For high pixel value input, it will be mapped to a position closer to the output pulse in the LTP window time; for zero input, it will be mapped to the LTD window. Since the emission time of the postsynaptic pulse is fixed in this scheme, the input with high pixel value is closer to the postsynaptic pulse, and in the typical STDP exponential decay learning curve, it can index to a larger Δw LTPThe zero input will weaken the synaptic connection of the position under the inhibition effect of LTD (the zero input must be processed by means of weight decay in the firing rate coding).

[0099] In an embodiment, referring to Figure 6 , the pulse pair is determined by the following steps:

[0100] If it is detected that any neuron in the brain-like synaptic learning circuit emits a pulse, a global inhibition signal is generated to inhibit the pulse emission activities of all other neurons in step 113.

[0101] The pulse pair is determined based on the inhibition amplitude corresponding to the global inhibition signal in step 114.

[0102] The circuit also adopts a side inhibition mechanism. After any neuron in the chip emits a pulse, a global inhibition signal is generated to inhibit the pulse emission activities of all other neurons, and the inhibition amplitude is adjustable in the chip. Without the side inhibition mechanism, all neurons will gradually emit pulses, which means that all neurons are participating in learning the current sample, which is harmful to the difference of the learning result. After the side inhibition mechanism is added, only a few neurons can "stand out" and continue to emit pulses, showing the tendency to the current sample, while other neurons avoid participating in the learning of the sample.

[0103] The concept of so-called homeostasis is also considered in the circuit. Homeostasis refers to universal normalization. The total sum of the weight values of neurons should be basically consistent during the learning process. After normalization, the synapses learned by STDP are still enhanced, and the synapses without pixel input should be weakened. However, in hardware, it is relatively difficult to implement division circuits, and the time sequence power consumption and other overheads are relatively large. In the design, a subtractor is used to weaken the synapses without any pulse number input. In the firing rate coding scheme, the zero input point of the four periphery cannot generate a pulse, and the synapse point cannot establish an STDP pulse time sequence dependence relationship due to the lack of any pulse input, so weight decay is needed to assist in training the weight.

[0104] Depending on the STDP learning algorithm, the coding scheme, the side inhibition, and the weight decay assistance mechanism as described above, the chip can achieve good unsupervised online learning. According to the learning results of the standard data set learned online on the hardware platform, it can be known that, depending on the full and fair competition learning (referring to side inhibition) of each neuron and the filtering of the edge signal (referring to weight decay), different neurons can learn different input samples.

[0105] If STDP is applied to PAICORE2.0 multi-core chips with 16 online cores, more abundant learning capacity and effect can be achieved, and even the lateral inhibition mechanism can be used as a kind of supervision signal to inhibit some specified cores so that they only participate in learning for some fixed inputs.

[0106] It should be noted that the implemented STDP is essentially a kind of unsupervised learning, so if classification is to be achieved, the trained neurons need to be assigned labels, and the experimental effect of STDP learning for some classic classification tasks, including the use of a neural network-based classifier, is assigned a class according to the number of neuron pulses in the inference mode, and the additional classifier is trained offline. It can be found that the network architecture has a significant impact on the classification accuracy. The inference results show that the classification accuracy of the STDP learning method proposed by the design for the MNIST, Fashion-MNIST, EMNIST, and Chinese-MNIST data sets is 97.62%, 84.98%, 88.74%, and 91.90%, respectively, in a two-layer fully connected 100 hidden neuron network.

[0107] Since the accuracy may depend to a large extent on the label assignment process, it may mask the actual learning effect. A more effective evaluation should fully consider the nature of STDP learning, i.e. reconstructing the input features. The synaptic weights after STDP learning can intuitively reflect the learning preferences of neurons for a series of images. To further evaluate to what extent the trained weights can represent the original images, the design also trains a single-layer network using the trained weights (as the training set). Then, the design inputs the original input images (as the test set) into the network for inference and collects the proportion of correct responses. The evaluation results include the reconstruction accuracy on the left, and the class-based confusion matrix is listed on the right. It is worth noting that the reconstruction accuracy can be regarded as an overall evaluation of the STDP learning to extract the most common input features. With the increase of the number of neurons, the feature memory capacity will also be improved, resulting in better performance on all data sets.

[0108] The following describes the neuromorphic hardware system of the brain-like technology provided by the application. The neuromorphic hardware system of the brain-like technology described below can be correspondingly referred to the brain-like synaptic learning method described above.

[0109] Please refer to Figure 7 The application provides a neuromorphic hardware system of brain-like technology, comprising:

[0110] The pulse input module 710 is configured to determine a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in the brain-like synaptic learning circuit, where the pulse pair includes an input pulse generated by the presynaptic neuron and an output pulse generated by the postsynaptic neuron.

[0111] The synaptic weight calculation module 720 is configured to determine, based on the pulse pair, an STDP mechanism corresponding to the pulse pair, and a synaptic weight corresponding to the STDP mechanism.

[0112] The brain-like synaptic learning module 730 is configured to perform, based on the pulse pair and the synaptic weight, STDP learning corresponding to the brain-like synaptic learning circuit.

[0113] The STDP mechanism is a pulse time-dependent plasticity mechanism, the brain-like synaptic learning circuit includes a synapse, the presynaptic neuron and the postsynaptic neuron, and the pulses in the pulse pair correspond to a long-term potentiation process or a long-term depression process in the STDP mechanism according to a time sequence.

[0114] Further, the synaptic weight calculation module is further configured to:

[0115] determine, based on the pulse pair, a nearest-neighbor peak of the pulse pair;

[0116] determine, based on the nearest-neighbor peak and a weight amplitude generated by the STDP mechanism, a synaptic weight corresponding to the STDP mechanism.

[0117] Further, the synaptic weight calculation module is further configured to:

[0118] determine, based on a time at which the pulse pair is generated, a time difference between the pulses in the pulse pair;

[0119] if the time difference is greater than or equal to a preset threshold, determine that the weight amplitude is generated by the long-term potentiation process;

[0120] if the time difference is less than the preset threshold, determine that the weight amplitude is generated by the long-term depression process.

[0121] Further, the brain-like synaptic learning circuit further includes a first input pulse determination module configured to:

[0122] obtain an input pulse sample;

[0123] perform Poisson distribution process weighting on the input pulse sample to obtain a preprocessed sample corresponding to the input pulse sample;

[0124] determine, based on a preset time window, a pulse number and a proportion of the preprocessed sample in the preset time window.

[0125] The Poisson distribution process weighting is used to weight the input pulse samples based on the weighting value of the Poisson distribution, and the preprocessed samples are used to characterize the input pulses in terms of the number of pulses and the proportion.

[0126] Furthermore, the neuromorphic synaptic learning circuit also includes a second input pulse determination module, used for:

[0127] Acquire input pulse samples;

[0128] The input pulse sample is time-encoded so that each pixel value of the input pulse sample corresponds to a pulse, and the magnitude of each pixel value of the input pulse sample is mapped to a specific moment in a preset time window to obtain the input pulse.

[0129] Furthermore, the neuromorphic synaptic learning circuit also includes a pulse pair determination module, used for:

[0130] If any neuron in the neuromorphic synaptic learning circuit is detected to fire a pulse, a global inhibition signal is generated, which is used to suppress the pulse firing activity of all other neurons.

[0131] The pulse pair is determined based on the suppression amplitude corresponding to the global suppression signal.

[0132] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logic instructions in the memory 830 to execute a brain-like synaptic learning method, which includes: determining a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in the brain-like synaptic learning circuit, the pulse pair including an input pulse generated by the presynaptic neuron and an output pulse generated by the postsynaptic neuron; determining, based on the pulse pair, the STDP mechanism corresponding to the pulse pair and the synaptic weights corresponding to the STDP mechanism; and performing STDP learning corresponding to the brain-like synaptic learning circuit based on the pulse pair and the synaptic weights; wherein, the STDP mechanism is a pulse time-dependent plasticity mechanism, the brain-like synaptic learning circuit includes a synapse, the presynaptic neuron and the postsynaptic neuron, and the preceding and following pulses of the pulse pair correspond to the long-term potentiation process or long-term inhibition process in the STDP mechanism according to their temporal order.

[0133] Further, the logic instructions in the memory 830 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the part of the technical solutions that make essential contributions or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0134] In another aspect, the present application also provides a computer program product, the computer program product comprising a computer program, the computer program being stored in a non-transitory computer readable storage medium, and the computer program being executable by a processor to cause a computer to perform the brain-like synapse learning method provided by the above-mentioned methods. The method comprises: determining a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in the brain-like synapse learning circuit, the pulse pair comprising an input pulse generated by the presynaptic neuron and an output pulse generated by the postsynaptic neuron; determining, based on the pulse pair, an STDP mechanism corresponding to the pulse pair, and a synaptic weight corresponding to the STDP mechanism; and performing, based on the pulse pair and the synaptic weight, an STDP learning corresponding to the brain-like synapse learning circuit. The STDP mechanism is a pulse time-dependent plasticity mechanism, the brain-like synapse learning circuit comprises a synapse, the presynaptic neuron and the postsynaptic neuron, and the pulses in the pulse pair correspond to a long-term potentiation process or a long-term depression process in the STDP mechanism according to the time sequence.

[0135] In yet another aspect, the application also provides a non-transitory computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the brain-like synapse learning method provided by any of the above methods, and the method comprises: determining a pulse pair generated by a presynaptic neuron and a postsynaptic neuron in the brain-like synapse learning circuit, the pulse pair comprising an input pulse generated by the presynaptic neuron and an output pulse generated by the postsynaptic neuron; determining, based on the pulse pair, an STDP mechanism corresponding to the pulse pair, and a synaptic weight corresponding to the STDP mechanism; and performing, based on the pulse pair and the synaptic weight, STDP learning corresponding to the brain-like synapse learning circuit; wherein the STDP mechanism is a pulse time-dependent plasticity mechanism, the brain-like synapse learning circuit comprises a synapse, the presynaptic neuron and the postsynaptic neuron, and the pulses in the pulse pair correspond to a long-term potentiation process or a long-term depression process in the STDP mechanism according to a time sequence.

[0136] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0137] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be implemented by means of software plus necessary general hardware platforms, and of course can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in terms of the contribution to the prior art, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0138] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A brain-like synaptic learning method, characterized in that, The method is applied to a neuromorphic hardware system based on neuromorphic technology, the neuromorphic hardware system including a neuromorphic synaptic learning circuit, the method comprising: Identify the pulse pairs generated by presynaptic neurons and postsynaptic neurons in the brain-like synaptic learning circuit, wherein the pulse pairs include input pulses generated by the presynaptic neurons and output pulses generated by the postsynaptic neurons; Based on the pulse pair, determine the STDP mechanism corresponding to the pulse pair, and the synaptic weights corresponding to the STDP mechanism; Based on the pulse pairs and the synaptic weights, STDP learning is performed on the neuromorphic synaptic learning circuit. The STDP mechanism is a pulse time-dependent plasticity mechanism. The brain-like synaptic learning circuit includes a synapse, a presynaptic neuron, and a postsynaptic neuron. The preceding and following pulses of the pulse pair correspond to the long-term enhancement process or long-term inhibition process in the STDP mechanism according to their temporal order. Based on the pulse pairs, the synaptic weights corresponding to the STDP mechanism are determined, including: Based on the pulse pair, determine the peak value of the nearest neighbor of the pulse pair; Based on the nearest neighbor peak value and the weight magnitude generated by the STDP mechanism, the synaptic weight corresponding to the STDP mechanism is determined; the weight magnitude generated by the STDP mechanism is determined based on the following steps: Based on the time when the pulse pair is generated, determine the time difference between the preceding and following pulses of the pulse pair; If the time difference is greater than or equal to a preset threshold, then the weight magnitude is determined to be generated by the long-term enhancement process; If the time difference is less than the preset threshold, then the weight magnitude is determined to be generated by the long-term suppression process; Based on the peak value of the nearest neighbor and the weight magnitude generated by the STDP mechanism, the synaptic weights corresponding to the STDP mechanism are determined, including: The nearest neighbor peak value and the weight magnitude generated by the STDP mechanism are summed to calculate the synaptic weight corresponding to the STDP mechanism; The weight amplitude is obtained based on a lookup table, where the index address of the lookup table is the time difference between the preceding and following pulses. The firing time of the preceding and following pulses is obtained based on the reading from the pre- and post-synaptic pulse counters. All pulse counters before and after the synapse are first reset to their maximum values. Whenever a pre- or post-synaptic pulse arrives, the pre- or post-synaptic pulse counter is cleared to zero. In the next time step, if no pulse arrives, the counter will increment by one each time until it reaches its maximum value.

2. The neuromorphic synaptic learning method according to claim 1, characterized in that, The input pulse is determined through the following steps: Acquire input pulse samples; The input pulse samples are weighted using a Poisson distribution process to obtain the preprocessed samples corresponding to the input pulse samples; Based on a preset time window, the number and proportion of pulses in the preprocessed sample within the preset time window are determined; The Poisson distribution process weighting is used to weight the input pulse samples based on the weighting value of the Poisson distribution, and the preprocessed samples are used to characterize the input pulses in terms of the number of pulses and the proportion.

3. The neuromorphic synaptic learning method according to claim 1, characterized in that, The input pulse is determined through the following steps: Acquire input pulse samples; The input pulse sample is time-encoded so that each pixel value of the input pulse sample corresponds to a pulse, and the magnitude of each pixel value of the input pulse sample is mapped to a specific moment in a preset time window to obtain the input pulse.

4. The neuromorphic synaptic learning method according to claim 3, characterized in that, The pulse pair is determined through the following steps: If any neuron in the neuromorphic synaptic learning circuit is detected to fire a pulse, a global inhibition signal is generated, which is used to suppress the pulse firing activity of all other neurons. The pulse pair is determined based on the suppression amplitude corresponding to the global suppression signal.

5. A neuromorphic hardware system based on brain-like technology, characterized in that, include: A pulse input module is used to determine the pulse pairs generated by presynaptic neurons and postsynaptic neurons in a neuromorphic synaptic learning circuit, wherein the pulse pairs include input pulses generated by the presynaptic neurons and output pulses generated by the postsynaptic neurons; The synaptic weight calculation module is used to determine the STDP mechanism corresponding to the pulse pair and the synaptic weight corresponding to the STDP mechanism based on the pulse pair. A neuromorphic synapse learning module is used to perform STDP learning corresponding to the neuromorphic synapse learning circuit based on the pulse pairs and the synaptic weights. The STDP mechanism is a pulse time-dependent plasticity mechanism. The brain-like synaptic learning circuit includes a synapse, a presynaptic neuron, and a postsynaptic neuron. The preceding and following pulses of the pulse pair correspond to the long-term enhancement process or long-term inhibition process in the STDP mechanism according to their temporal order. Based on the pulse pairs, the synaptic weights corresponding to the STDP mechanism are determined, including: Based on the pulse pair, determine the peak value of the nearest neighbor of the pulse pair; Based on the nearest neighbor peak value and the weight magnitude generated by the STDP mechanism, the synaptic weight corresponding to the STDP mechanism is determined; the weight magnitude generated by the STDP mechanism is determined based on the following steps: Based on the time when the pulse pair is generated, determine the time difference between the preceding and following pulses of the pulse pair; If the time difference is greater than or equal to a preset threshold, then the weight magnitude is determined to be generated by the long-term enhancement process; If the time difference is less than the preset threshold, then the weight magnitude is determined to be generated by the long-term suppression process; Based on the peak value of the nearest neighbor and the weight magnitude generated by the STDP mechanism, the synaptic weights corresponding to the STDP mechanism are determined, including: The nearest neighbor peak value and the weight magnitude generated by the STDP mechanism are summed to calculate the synaptic weight corresponding to the STDP mechanism; The weight amplitude is obtained based on a lookup table, where the index address of the lookup table is the time difference between the preceding and following pulses. The firing time of the preceding and following pulses is obtained based on the reading from the pre- and post-synaptic pulse counters. All pulse counters before and after the synapse are first reset to their maximum values. Whenever a pre- or post-synaptic pulse arrives, the pre- or post-synaptic pulse counter is cleared to zero. In the next time step, if no pulse arrives, the counter will increment by one each time until it reaches its maximum value.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the neuromorphic synaptic learning method as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the neuromorphic synaptic learning method as described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the neuromorphic synaptic learning method as described in any one of claims 1 to 4.

Images