Learning algorithm of brain-like expression and calculation unit model

Abstract

The invention provides a brain-like expression and calculation unit learning algorithm, and the brain-like expression and calculation unit comprises an input layer, a compression expression layer, a sparse expression layer, and a self-organization mapping expression layer. The learning algorithm of the brain-like expression and calculation unit comprises three parts, the first part is learning of a compression expression layer and a sparse expression layer, the second part is learning of a self-organizing mapping expression layer, and the third part is synchronous oscillation learning among different self-organizing mapping expression layers; the learning of the compression expression layer and the sparse expression layer comprises three learning parameters, namely the weight of the compression expression layer and the dictionary and sparse vector of the sparse expression layer, and the objective function is shown in the specification; learning of the self-organizing mapping expression layer is updated according to a pulse timing dependence plasticity rule; synchronous oscillation learning between different self-organizing mapping expression layers is realized through dynamic connection strength training between self-organizing networks.

Description

Technical Field

[0001] The present application relates to brain-like model technology, and in particular to a learning algorithm for a brain-like expression and computing unit model. Background Art

[0002] Existing brain-like models are mostly based on Transformer, a deep learning architecture characterized by the use of an attention mechanism, which allows the model to assign different weights to each element in the input sequence and capture dependencies in the sequence. Training methods can include semi-supervised learning, online learning, and reinforcement learning, but the data and computing resources required are relatively high. Currently, the training of large-scale spiking neural networks mostly uses the ANN-SNN method, which is inconsistent with the actual working mechanism of the brain and the working mechanism of the human brain, and lacks biological rationality and solvability. Summary of the invention

[0003] In view of the above problems, the present application proposes a learning algorithm based on a brain-like expression and computing unit model, which can simulate the biological mechanism of sensory encoding in the human brain and increase its biological rationality and interpretability.

[0004] The learning algorithm of the brain-like expression and computing unit of the present application, wherein: The brain-like expression and calculation unit includes: an input layer, a compression expression layer, a sparse expression layer and a self-organizing map expression layer; The input layer is used to receive raw data or information; The compression expression layer is used to receive the data from the input layer and the output of the compression expression layer of this unit at the previous moment; The sparse expression layer is used for sparse coding, receiving the data of the compressed expression layer, independently sparsely encoding the data and outputting it to the self-organizing map expression layer; The self-organizing map expression layer is composed of a self-organizing map network; the output data of the self-organizing map expression layer is stretched into a self-organizing map expression space; the self-organizing map expression layer projects the output data of the compression expression layer or the sparse expression layer into the self-organizing map expression space; The learning algorithm of the brain-like expression and computing unit includes three parts, wherein the first part is the learning of the compressed expression layer and the sparse expression layer, the second part is the learning of the self-organizing map expression layer, and the third part is the synchronous oscillation learning between different self-organizing map expression layers; The learning of the compressed expression layer and the sparse expression layer includes three learning parameters, namely, the weight of the compressed expression layer , dictionary of sparse representation layer and sparse vectors , the objective function is ; The learning of the expression layer of the self-organizing map is updated according to the spike timing-dependent plasticity rule; Synchronous oscillation learning between different self-organizing map expression layers is achieved by training the dynamic connection strengths between self-organizing networks.

[0005] Preferably, in the learning of the compressed expression layer and the sparse expression layer, the original data or information The compressed expression vector is obtained through the compressed expression layer, and then through the sparse expression layer to learn the dictionary and sparse vectors , calculate the sparse expression error , taking this as the error of the compressed expression layer, and updating the weights through the back propagation algorithm .

[0006] Preferably, the original data or information The compressed expression vector is obtained by compressing the expression layer, and the preset dictionary is used and sparse vectors Calculate the reconstruction vector and calculate the loss function , taking this as the error of the compressed expression layer, and updating the weights through the back propagation algorithm , again Input the compressed expression layer to obtain the compressed expression vector, and then pass it through the sparse expression layer to learn the dictionary and sparse vectors .

[0007] Preferably, the input of the compressed representation layer is the original data or information Multi-channel data after various data enhancement methods; loss function It is the sum of pairwise errors between sparsely reconstructed vectors of different channels.

[0008] The learning algorithm of the brain-like expression and computing unit proposed in the present invention simulates the information processing mechanism of human sensory organs and learns according to the biological mechanism of brain sensory encoding. It has strong biological rationality and explainability. The model can be trained by unsupervised training methods, and the input data does not need to be labeled separately, laying the foundation for the development of large-scale brain-like expression and computing models. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a schematic diagram of the architecture of the brain-like expression and computing unit model involved in this application; Figure 2 for Figure 1 Schematic diagram of parameters related to the compressed expression layer and sparse expression layer of the brain-like expression and computing unit model; Figure 3 for Figure 1A vibration model of reentrant connections coupled between self-organizing expression layers of a brain-like expression and computational unit model. DETAILED DESCRIPTION

[0010] The learning algorithm of the brain-like expression and computing unit of this application is based on Figure 1 The brain-like architecture shown.

[0011] This type of brain architecture includes: input layer, compression expression layer, sparse expression layer, self-organizing map expression layer and compression coding succession layer.

[0012] Input layer: used to receive raw data or information.

[0013] Compression expression layer: receives data from the input layer, the output of the compression expression layer of this unit at the previous moment, and the output of the compression expression layer of other neighboring units.

[0014] Compression coding receiving layer: stores the output of the compression expression layer of each unit, which is used for splicing and fusion between units and system integration.

[0015] Sparse expression layer: used for sparse coding, receiving data from the compressed expression layer, which is independently sparsely encoded and output to the self-organizing map expression layer.

[0016] Self-organizing map expression layer: It is composed of a self-organizing map network and receives output data from the compressed expression layer or the sparse expression layer.

[0017] The output data of the self-organizing map expression layer is stretched into a self-organizing map expression space; the self-organizing map expression layer projects the output data of the compression expression layer or the sparse expression layer into the self-organizing map expression space.

[0018] The learning algorithm of the brain-like expression and computing unit of the present application is divided into three parts. The first part is the learning of the compressed expression layer and the sparse expression layer, the second part is the learning of the self-organizing map expression layer, and the third part is the synchronous oscillation learning between different self-organizing map expression layers.

[0019] Part I: Learning of compressed representation layer and sparse representation layer, such as Figure 2 shown.

[0020] There are three learning parameters in this part, namely the weight of the compressed expression layer , dictionary of sparse representation layer and sparse vectors , the objective function is ,Since the objective function is an extremely complex nonlinear function, we optimize it in three steps. Dictionary with sparse representation layer and sparse vectors The learning can be done by first presetting the weights of the compressed expression layer Let's first learn the dictionary of the sparse representation layer and sparse vectors Specifically, the data The compressed expression vector is obtained through the compressed expression layer, and then through the sparse expression layer to learn the dictionary and sparse vectors , calculate the sparse expression error , taking this as the error of the compressed expression layer, and updating the weights through the back propagation algorithm , complete the first part of the study. You can also preset the dictionary of the sparse expression layer first and sparse vectors To learn the weights of the compressed representation layer Specifically, the data The compressed expression vector is obtained by compressing the expression layer, and the preset dictionary is used and sparse vectors Calculate the reconstruction vector and calculate the loss function , taking this as the error of the compressed expression layer, and updating the weights through the back propagation algorithm , again Input the compressed expression layer to obtain the compressed expression vector, and then pass it through the sparse expression layer to learn the dictionary and sparse vectors , and so on, to complete the first part of the learning. Among them, the input of the compressed expression layer can also be multi-channel data after the original data has been enhanced by various data methods. In this case, the loss function It is the sum of pairwise errors between sparsely reconstructed vectors of different channels.

[0021] Part II: Learning the Self-Organizing Map Representation Layer The SOM expression layer is updated according to the spike timing dependent plasticity rule (STDP). The spike timing dependent plasticity rule is a biological process that regulates the strength of connections between neurons in the brain. This process adjusts the connection strength according to the relative timing of the output and input action potentials (or spikes) of a particular neuron. When the presynaptic neuron fires a spike before the postsynaptic neuron, the weight between the two increases. Conversely, when the postsynaptic neuron fires a spike before the presynaptic neuron, the weight between the two decreases. If the output signal of the compressed expression layer is an analog signal, it needs to be converted into a spike train and then input into the SOM expression layer. This method is applicable to both SOMs of compressed expression signals and SOMs of sparsely coded expression signals.

[0022] Part III: Learning of reentrant oscillations between different SOM representation layers The network training of reciprocating connection weights between the oscillations and synchronous oscillations between neurons in the self-organizing map expression layer. This symbolic brain-like oscillation and synchronous oscillation form a connection network diagram, which realizes global communication, association and task bundling of brain connection diagrams of different tasks, such as Figure 3 shown. Figure 3 It shows how to achieve oscillations and synchronized oscillations between brain area networks and cortical networks, which are achieved by training the dynamic connection strength between self-organizing networks.

[0023] The synchronized oscillation coupling signals output by the self-organizing map expression layer can then be fed into the controller to perform tasks such as behavioral and cognitive control.

[0024] Solution 1: The input data is image data, and the input data is I Model initialization: compression coding layer weight W, regularization parameter , update the number of iterations N, the time length of the pulse sequence T, and the weight adjustment parameter and , the neuron time constant and .

[0025] Part 1: Learning Compressed and Sparse Representation Layers For n=1 to N: 1. Input Data In I, after the compressed expression layer, the compressed expression vector is obtained ; 2. Update the dictionary in two steps and ; Step 1: Dictionary learning and sparse coding; (1) Initialize the dictionary , is a randomly generated matrix; (2) Initialize the sparse coefficient matrix is an all-zero matrix; (3) Initialize the matrix , ; (4) for t = 1 to T: a. From Select data ; b. Solve the coefficient sparse matrix ; Solve using the least squares method ; c. Update ; d. Update ; e. Calculate using step2 algorithm ; f. End for; (5) Return , Step 2: Dictionary update; Input: dictionary matrix ,matrix , matrix ; (1) For j = 1 to k: Update according to the following formula : u j ← 1 A [ j , j ] ( b j - D a j )+ d j ; ; (2) End for; (3) Return to D; 3. Update the weights and biases of the compressed expression layer according to the back-propagation algorithm; For each training example: Step 1: forward propagation, calculating error; (1) Reconstructing the compressed expression vector (2) Calculate the reconstruction error Loss = F( ); Step 2: Backward propagation: Loss.backward; Step 3: Update weights and biases: Optimizer.step; 4.End for; Part II: Learning of the Self-Organizing Map Representation Layer; 1. The output α of the sparse expression layer is pulsed and converted into a pulse train with a time length of T, which is sent to the self-organizing map expression layer; 2.For t=1 to T: Step 1: Calculate the output pulses of the input layer neurons; The pulsed sparse matrix is ​​sent to the input layer neurons (taking LIF neurons as an example), and the output pulses of each neuron in the input layer are calculated according to the following formula, and the time when each neuron first emits a pulse is recorded, which is recorded as ; ; ; ; Step 2: Calculate the output pulses of the competition layer neurons; The pulse sequence output by the input layer is sent to the competitive layer neurons (LIF neurons, see step 1 for the formula), the output pulses of the competitive layer neurons are calculated, and the time when each neuron first emits a pulse is recorded, which is recorded as ; Step 3: Update weights; (1) Calculation ,in ; ; (2) Update W; 3.End for; Part III: Oscillations of reentrant connections between layers of the self-organizing map expression; Taking the self-organizing map expression of sparse expression signals as an example, the same is true for compressed expression signals; 1. The output pulses of two different self-organizing map expression layers are recorded as and , comes from the input connection weight of the sparse representation layer and , i and j represent neurons from two different self-organizing map expression layers, respectively, and T is the time length of the spike train; 2. For t=1 to T: Step 1: Calculate the activity of the i-th and j-th neurons; ; ; in and Represents the lateral connection weights of the neighboring neurons of the i-th and j-th neurons; Step 2: Calculate the mean field ; ; and ; Step 3: Calculate the dynamic connection weight between neuron i and neuron j; ; 3. End for; Solution 2: The input data is image data, and the input data is I; Model initialization: compression coding layer weight W, dictionary , a sparse vector , the regularization parameter , update the number of iterations N, the time length of the pulse sequence T, and the weight adjustment parameter and , the neuron time constant and .

[0026] Part 1: Learning of compressed expression layer and sparse expression layer; For n=1 to N: 1. Input Data In I, after the compressed expression layer, the compressed expression vector is obtained ; 2. Update the weights and biases of the compressed representation layer according to the back-propagation algorithm; Step 1: forward propagation, calculating error; (1) Reconstructing the compressed expression vector ; (2) Calculate the reconstruction error Loss = F( ); Step 2: Backward propagation: Loss.backward; Step 3: Update weights and biases: Optimizer.step; 3. Calculate the updated compressed expression vector: ; 4. Update the dictionary in two steps and ; Step 1: Dictionary learning and sparse coding; (1) Initialize the matrix , (2) for t=1 to T: a. From Select data ; b. Solve the coefficient sparse matrix Solve using the least squares method ; c. Update ; d. Update ; e. Calculate using step2 algorithm (3) End for; (4) Return , ; 5. End for; Step 2: Dictionary update; Input: dictionary matrix ,matrix , matrix ; (1) For j = 1 to k; Update according to the following formula ; u j ← 1 A [ j , j ] ( b j - D a j )+ d j ; ; (2) End for; (3) Return to D; Part II: Learning of the Self-Organizing Map Representation Layer; 1. The output α of the sparse expression layer is pulsed and converted into a pulse train with a time length of T, which is sent to the self-organizing map expression layer; 2.For t=1 to T: Step 1: Calculate the output pulses of the input layer neurons; The pulsed sparse matrix is ​​sent to the input layer neurons (taking LIF neurons as an example), and the output pulses of each neuron in the input layer are calculated according to the following formula, and the time when each neuron first emits a pulse is recorded, which is recorded as ; ; ; ; Step 2: Calculate the output pulses of the competition layer neurons; The pulse sequence output by the input layer is sent to the competitive layer neurons (LIF neurons, see step 1 for the formula), the output pulses of the competitive layer neurons are calculated, and the time when each neuron first emits a pulse is recorded, which is recorded as Step 3: Update weights; (1) Calculation ,in ; ; (2) Update W; 3.End for; Part III: Oscillations of reentrant connections between layers of the self-organizing map expression; Taking the self-organizing map expression of sparse expression signals as an example, the same is true for compressed expression signals; 1. The output pulses of two different self-organizing map expression layers are recorded as and , comes from the input connection weight of the sparse representation layer and , i and j represent neurons from two different self-organizing map expression layers, respectively, and T is the time length of the spike train; 2. For t=1 to T: Step 1: Calculate the activity of the i-th and j-th neurons; ; ; in and Represents the lateral connection weights of the neighboring neurons of the i-th and j-th neurons; Step 2: Calculate the mean field ; ; and ; Step 3: Calculate the dynamic connection weight between neuron i and neuron j; ; 3. End for; Solution 3: The input data is image data, and the input data is I; Model initialization: compression coding layer weight W, regularization parameter , update the number of iterations N, the time length of the pulse sequence T, and the weight adjustment parameter and , the neuron time constant and .

[0027] Part 1: Learning of compressed expression layer and sparse expression layer; For n=1 to N: 1. Input Data In I, suppose that after 3 different data enhancements, the data is obtained , , , after the compressed expression layer, the compressed expression vector is obtained , , ; 2. Update the dictionary in two steps and ; Step 1: Dictionary learning and sparse coding; (1) Initialize the dictionary , is a randomly generated matrix; (2) Initialize the sparse coefficient matrix is an all-zero matrix; (3) Initialize the matrix , ; (4) for t = 1 to T: g. From Select data ; h. Solve the coefficient sparse matrix ; Solve using the least squares method ; i. Update ; j. Update ; k. Calculate using step2 algorithm ; l. End for; (5) Return , ; (6) Repeat (4)-(5) and calculate and D and α afterwards; Step 2: Dictionary update; Input: dictionary matrix ,matrix , matrix ; (1) For j = 1 to k: Update according to the following formula : u j ← 1 A [ j , j ] ( b j - D a j )+ d j ; ; (2) End for; (3) Return to D; 3. Update the weights and biases of the compressed expression layer according to the back-propagation algorithm; For each training example: Step 1: forward propagation, calculating error; (3) According to the formula Reconstruct the compressed expression vector and get , , ; (4) Calculate the reconstruction error Loss = F( )+ F( )+ F( ); Step 2: Backward propagation: Loss.backward; Step 3: Update weights and biases: Optimizer.step; 4.End for; Part II: Learning of the Self-Organizing Map Representation Layer; 1. The output α of the sparse expression layer is pulsed and converted into a pulse train with a time length of T, which is sent to the self-organizing map expression layer; 2.For t=1 to T: Step 1: Calculate the output pulses of the input layer neurons; The pulsed sparse matrix is ​​sent to the input layer neurons (taking LIF neurons as an example), and the output pulses of each neuron in the input layer are calculated according to the following formula, and the time when each neuron first emits a pulse is recorded, which is recorded as ; ; ; ; Step 2: Calculate the output pulses of the competition layer neurons; The pulse sequence output by the input layer is sent to the competitive layer neurons (LIF neurons, see step 1 for the formula), the output pulses of the competitive layer neurons are calculated, and the time when each neuron first emits a pulse is recorded, which is recorded as ; Step 3: Update weights; (1) Calculation ,in ; ; (2) Update W; 3.End for; Part III: Oscillations of reentrant connections between layers of the self-organizing map expression; Taking the self-organizing map expression of sparse expression signals as an example, the same is true for compressed expression signals; 1. The output pulses of two different self-organizing map expression layers are recorded as and , comes from the input connection weight of the sparse representation layer and , i and j represent neurons from two different self-organizing map expression layers, respectively, and T is the time length of the spike train; 2. For t=1 to T: Step 1: Calculate the activity of the i-th and j-th neurons; ; ; in and Represents the lateral connection weights of the neighboring neurons of the i-th and j-th neurons; Step 2: Calculate the mean field ; ; and ; Step 3: Calculate the dynamic connection weight between neuron i and neuron j; ; 3. End for.

Claims

1. A learning algorithm for brain-like expression and computational units, wherein: The brain-like expression and calculation unit includes: an input layer, a compression expression layer, a sparse expression layer and a self-organizing map expression layer; The learning algorithm of the brain-like expression and computing unit includes three parts, wherein the first part is the learning of the compressed expression layer and the sparse expression layer, the second part is the learning of the self-organizing map expression layer, and the third part is the synchronous oscillation learning between different self-organizing map expression layers; The learning of the compressed expression layer and the sparse expression layer includes three learning parameters, namely, the weight of the compressed expression layer , dictionary of sparse representation layer and sparse vectors , the objective function is ; The learning of the expression layer of the self-organizing map is updated according to the spike timing-dependent plasticity rule; Synchronous oscillation learning between different self-organizing map expression layers is achieved by training the dynamic connection strengths between self-organizing networks.

2. The learning algorithm of the brain-like expression and calculation unit according to claim 1, characterized in that: In the learning of compressed expression layer and sparse expression layer, the original data or information The compressed expression vector is obtained through the compressed expression layer, and then through the sparse expression layer to learn the dictionary and sparse vectors , calculate the sparse expression error , taking this as the error of the compressed expression layer, and updating the weights through the back propagation algorithm .

3. The learning algorithm of the brain-like expression and calculation unit according to claim 1, characterized in that: Raw data or information The compressed expression vector is obtained by compressing the expression layer, and the preset dictionary is used and sparse vectors Calculate the reconstruction vector and calculate the loss function , taking this as the error of the compressed expression layer, and updating the weights through the back propagation algorithm , again Input the compressed expression layer to obtain the compressed expression vector, and then pass it through the sparse expression layer to learn the dictionary and sparse vectors .

4. The learning algorithm of the brain-like expression and calculation unit according to claim 2 or 3, characterized in that: The input of the compressed representation layer is the original data or information Multi-channel data after various data enhancement methods; loss function It is the sum of pairwise errors between sparsely reconstructed vectors of different channels.

Images