A continuous learning control network for cerebellar dynamic sequence motion and its method

Abstract

The application discloses a cerebellar dynamic sequence motion continuous learning control network and a method thereof. The control network comprises an input layer, an intermediate layer and an output layer. The intermediate layer is composed of a feedforward inhibition module and a feedback inhibition module. The input layer generates a mossy fiber signal with dynamic timing. The intermediate layer sparsely samples the input mossy fiber signal through granule cells, and the inhibition module dynamically modulates the activity level and encoding mode of the granule cells. The feedforward inhibition module averages the mossy fiber signal discharge rate to obtain the activity attribute of Golgi cells, which is inherited to the granule cells. The feedback inhibition module sparsely samples the granule cells to obtain the activity attribute of Golgi cells, which is inherited to the granule cells. The output layer sequentially outputs the multiplication result of the activity range of the granule cells at each time point and the connection weight between the corresponding Purkinje cells as the control cerebellar signal. The application improves the problem that the biological constraint cerebellar network is not accurate and the motion conversion is not smooth in sequence motion continuous learning.

Description

Technical Field

[0001] This invention relates to the field of brain-like intelligence technology, and in particular to a continuous learning control network and method for cerebellar dynamic sequence motion. Background Technology

[0002] Sequential motor behavior in organisms is composed of a series of basic motor modules flexibly combined at different times, which is of great significance to individual survival and evolution. The cerebellum receives and integrates afferent information from the cerebral cortex, spinal cord, and sensory system, and participates in the learning, optimization, and execution conversion of complex motor sequences in real time to improve the accuracy, fluency, and coordination of different motor modules. Compared with the cerebrum, the cerebellum's circuit structure is relatively simple but highly homogeneous and ordered. The main information transmission path is usually equivalent to a three-layer feedforward neural network. Traditional cerebellar models in sensorimotor learning theory and research often focus on the learning of individual motor paradigms. However, the neural coding and learning mechanisms by which this simple circuit supports high-level motor functions such as accurate learning, robust execution, and flexible adjustment of motor sequences remain unclear and require further exploration and understanding.

[0003] Researchers have discovered that inhibitory neurons, specifically Golgi cells, in the cerebellar granule cell layer play a crucial role in regulating granule cell activity, enriching information encoding mechanisms, and flexibly transforming information. While coordinating granule cell activity to maintain network stability, they enhance input selectivity to improve the accuracy of information processing, ultimately increasing the cerebellum's efficiency in learning motor sequences. Inspired by this phenomenon, we explored and improved the cerebellum's learning ability in sequential movement by dynamically altering the encoding level and mode of granule cells by introducing two different Golgi cell-mediated inhibitory modules into the cerebellar granule cell layer: feedforward inhibition and feedback inhibition. Linking low-level cerebellar circuits with high-level motor sequence execution functions not only helps to deepen our understanding and explore the cerebellum's encoding mechanisms and computational advantages in motor learning, providing a theoretical basis for the understanding and development of cerebellar functional reuse and brain-like intelligence technologies, but also has significant meaning and reference value for thoroughly understanding and revealing the pathological mechanisms of motor regulation and movement disorders. Summary of the Invention

[0004] To address the challenges of existing technologies, this invention aims to provide a dynamic sequence motion learning and control network and method for the cerebellum based on the main neuron types, electrophysiological characteristics, and connection patterns between neurons in the cerebellar circuit. This enables efficient learning and execution of dynamic sequence motion within a biologically constrained cerebellar neural computational network model.

[0005] To address the problems of the existing technology, the present invention adopts the following technical solution:

[0006] A continuous learning control network for cerebellar dynamic sequence motion is proposed, comprising an input layer, an intermediate layer, and an output layer; the intermediate layer consists of a feedforward inhibition module and a feedback inhibition module.

[0007] Cheng; among which:

[0008] The input layer generates moss fiber signals with dynamic temporal sequences;

[0009] The intermediate layer sparsely samples the moss fiber signal input through granule cells, while the inhibition module dynamically modulates the granule cell activity level and encoding mode; wherein:

[0010] The feedforward inhibition module averages the signal discharge rate of moss fibers to obtain the activity attributes of Golgi cells, which are then inherited by granular cells.

[0011] The feedback inhibition module obtains the activity attributes of Golgi cells by sparse sampling of granulocytes and inherits them to granulocytes.

[0012] The output layer processes the granule cell activity and the corresponding Purkinje cells at each time point according to the following formula to control the output signal of the cerebellar Purkinje cells:

[0013] PC m (t)=∑W mj GC j (t)m=1,2,...,N3

[0014] Where: N3 represents the total number of Purkinje cells, GC j (t) represents the firing rate of the j-th granule cell in the intermediate layer, W mj This represents the connection weight between the j-th granule cell and the m-th Purkinje cell.

[0015] Furthermore, the intermediate layer sparsely samples the moss fiber signal input through granule cells, while simultaneously inhibiting the module's dynamic modulation of granule cell activity levels and encoding methods, including:

[0016] M1 moss fiber signals were randomly sampled from each granulocyte using the following formula and weighted accordingly. Uniform weighting was used to obtain the granulocyte firing rate:

[0017]

[0018] Where: f is the piecewise linear activation function f(x) = max(x, 0); MF i (t) represents the discharge rate of the i-th moss fiber in the input layer; M1 represents the number of moss fibers sampled for each granule cell; N1 represents the total number of granule cells.

[0019] Furthermore, the feedforward inhibition adjustment particle module obtains the activity attributes of Golgi cells and inherits them to granule cells by averaging the moss fiber signal discharge rate using the following formula:

[0020]

[0021] Where: θ represents the feedforward inhibition experienced by the j-th granule cell; λ1 represents the feedforward inhibition coupling strength. Further, the process by which the feedback inhibition granule module inherits the activity attributes of Golgi cells from sparse sampling of granule cells to granule cells includes:

[0022] M² granulocytes are randomly sampled from each Golgi cell using the following formula and weighted accordingly. Uniform weighting:

[0023]

[0024] Where: M2 represents the number of granular cells sampled per Golgi cell; N2 represents the total number of Golgi cells;

[0025] M3 Golgi cells were randomly sampled from each granulocyte using the following formula and weighted accordingly. Uniform weighting:

[0026]

[0027] Where: θ j (t) represents the feedback inhibition experienced by the j-th granule cell, and λ2 represents the feedback inhibition coupling strength.

[0028] To solve the technical problem, the present invention can also adopt the following technical solution:

[0029] S1. Construct a cerebellar dynamic sequence motor control network based on firing rate;

[0030] S2, Sparse sampling of intermediate layer granular cells: The granular cells are sparsely and uniformly received by the moss fiber input in the input layer according to the time sequence.

[0031]

[0032] Where f is the piecewise linear activation function f(x) = max(x,0); N1 represents the total number of granule cells;

[0033] M1 represents the number of moss fibers connected to each granule cell; MF i (t)

[0034] This represents the discharge rate of the i-th moss fiber in the input layer;

[0035] S3, Feedforward Inhibition Module Granulocyte Input: Golgi cells receive excitatory input from moss fibers in an averaged manner to form feedforward inhibition, which is then multiplied by a scaling parameter to dynamically adjust the magnitude of the granulocyte input and the corresponding activation level; that is:

[0036]

[0037] Where: θ represents the feedforward inhibition experienced by the j-th granule cell; λ1 represents the feedforward inhibition coupling strength;

[0038] S4, Feedback Inhibition Module Granulocyte Input: Golgi cells sparsely sample granulocytes and sparsely act on them to form feedback inhibition. This feedback inhibition is dynamically adjusted by multiplying the input to the granulocytes by a scaling parameter, thereby adjusting the size and corresponding activation level of the input.

[0039]

[0040] GoC k (k)=0k=1,2...,or N2

[0041]

[0042] Where: N1 represents the total number of granulocytes; N2 represents the total number of Golgi cells; GoC k (t) represents the firing rate of the kth Golgi cell in the intermediate layer; θ j (t) represents the feedback inhibition experienced by the j-th granule cell; λ2 represents the feedback inhibition coupling strength;

[0043] S5. Output layer Purkinje cells complete the task output: Multiply the granule cell activity after the inhibition module's action and the connection weights between Purkinje cells to obtain the final output required for the task; that is:

[0044] PC m (t)=ΣW mj GC j (t)j=1,2,...,N3

[0045] Where: N3 represents the total number of Purkinje cells; W mj This represents the connection weight between the j-th granule cell and the m-th Purkinje cell;

[0046] S6. Calculate the loss value and gradient descent: Statistically analyze the total loss value and gradient descent of the Purkinje cells in the output layer at each time point.

[0047] The output is comprehensively calculated and compared with the actual target value to calculate the loss value; that is:

[0048]

[0049] Where P(t) is a vector composed of the outputs of 3 Purkinje neurons, and T(t) is the expected 3D spatial trajectory vector;

[0050] The gradient descent learning algorithm is used to update the connection weights between granule cells and Purkinje cells; that is:

[0051] ΔW=W-(η*Err(t)*GC(t)′)

[0052] S7. Test network accuracy: If the network accuracy is stable or the maximum number of training iterations is reached, stop network training.

[0053] Beneficial effects

[0054] Compared with traditional technical solutions, the beneficial effects of this invention are:

[0055] This invention flexibly adjusts the encoding level and spatiotemporal characteristics of granule cells through a Golgi cell inhibition module. Based on a network structure using the Golgi cell inhibition module, this invention efficiently encodes and processes sensory input information, improving the cerebellar network's ability to overcome the inaccuracies and disjointed action transitions in continuous sequence motor learning under biological constraints. While improving learning accuracy, it also reduces noise interference along the time direction, enabling the cerebellar network to efficiently complete the learning and execution of sequential movements. Attached Figure Description

[0056] Figure 1 This invention provides a structural diagram of a dynamic sequence motion control network for cerebellar neurons.

[0057] Figure 2 This invention relates to a diagram illustrating the information processing of different inhibition modules in a dynamic sequence motion control network of cerebellar neurons.

[0058] Figure 3 It is based on the information encoding mechanism of the feedforward inhibition module and the display of single motion learning results.

[0059] Figure 4 It is based on the information encoding mechanism of the feedback inhibition module and the display of individual motion learning results.

[0060] Figure 5 This is a demonstration of the sequence motion learning results of a cerebellar network based on a feedback inhibition module. Detailed Implementation

[0061] The following is in conjunction with the appendix Figure 1 The present invention is described as follows:

[0062] This invention is primarily verified using simulation experiments; all steps and conclusions have been verified correctly using Matlab 2021b. The following is a summary of the appendix. Figure 1-5The embodiments of the present invention are described in detail below to enable those skilled in the art to better understand the present invention.

[0063] like Figure 1 As shown, the cerebellar neural circuit network structure based on the inhibition module involved in this invention consists of an input layer (containing moss fibers), an intermediate layer (containing granule cells and Golgi cells), and an output layer (containing Purkinje cells). The intermediate layer includes a first cell unit, a second cell unit, and an inhibition module. The number of cells in different layers is set according to specific tasks and needs, as follows:

[0064] 1. Input layer, i.e., moss fiber layer: The data that this invention can process includes dynamic temporal signals containing time information. In this embodiment of the invention, the motor task of handwritten digit transcription is used as an example for illustration, and the speech data is input into the network as activated moss fibers in chronological order.

[0065] 2. Intermediate layer, i.e., granule cell layer: This layer primarily extracts moss fiber input through sparse sampling of granule cells. Simultaneously, it dynamically adjusts the granule cell activity level and encoding method based on feedforward and feedback inhibition modules, such as... Figure 2 As shown, different inhibition modules have different information processing mechanisms. The feedforward inhibition module receives input directly from moss fibers by Golgi cells and exerts an inhibitory effect on granule cells. The feedback inhibition module receives input from granule cells by Golgi cells and affects the activity of granule cells through inhibition.

[0066] 3. Output layer, i.e., Purkinje cell layer: This layer, along with the granule cell layer, adopts a fully connected structure. The output of the Purkinje cells at each time step is statistically analyzed, and the results are jointly determined. In this embodiment of the invention, the output of each cell in the Purkinje cell layer represents a different dimension of the motion trajectory.

[0067] In a specific embodiment of the present invention, the training process of the fine cerebellar network model based on the inhibition module is as follows:

[0068] S1: Construct a three-layer cerebellar network based on firing rate. The number of cells in the input, intermediate, and output layers is selected according to the specific task, and corresponding inhibition modules are chosen. In one embodiment of this invention, for the handwritten digit speech transcription task of the TI-46 spokenword corpus speech dataset, a cerebellar network is constructed consisting of a moss fiber layer with 12 neurons, a granular cell layer with 50,000 granule cells and 5,000 Golgi cells, and a Purkinje cell layer with 3 neurons. The three neurons in the Purkinje cell layer (output layer) represent different dimensions x, y, and z of the three-dimensional motion trajectory. The epoch is set to 1000, and the learning rate is initialized to 0.007.

[0069] S2: Sparse and uniform sampling of the granular cell layer. In one embodiment of the present invention, for the handwritten digit transcription task of the TI-46spoken wordcorpus speech dataset, the 12-dimensional input data is input into 12 moss fibers in the moss fiber layer in chronological order. M1 moss fibers are randomly sampled from each granular cell and weighted accordingly. Each input is weighted evenly.

[0070]

[0071] Where f is a piecewise linear activation function f(x) = max(x,0), M1 represents the number of moss fibers connected to each granule cell; MF i (t) represents the discharge rate of the i-th moss fiber in the input layer; GC j (t) represents the discharge rate of the j-th granule cell in the intermediate layer;

[0072] S3: Feedforward inhibition adjusts granulocyte input. Feedforward inhibition, generated by Golgi cells receiving input directly from excitatory moss fibers, is multiplied by scaling parameters to dynamically adjust the spatiotemporal characteristics of granulocytes and the corresponding granulocyte activation levels, such as... Figure 3 As shown. In some embodiments of the present invention, for moss fiber input, the average discharge rate of all moss fibers at time t is used as the activity level of Golgi cells and acts on the input inheritance process of granule cells, as shown in the following formula:

[0073]

[0074] Where: N1 represents the total number of granule cells; θ represents the feedforward inhibition experienced by the j-th granule cell; λ1 represents the feedforward inhibition coupling strength;

[0075] S4: Feedback inhibition adjusts granulosa cell input. Input received from granulosa cells by Golgi cells is influenced by inhibitory mechanisms, generating feedback inhibition. This inhibition, multiplied by scaling parameters, dynamically adjusts the spatiotemporal characteristics of granulosa cells and their corresponding activation levels. Figure 4 As shown. In some embodiments of the present invention, M2 granule cells are randomly sampled from each Golgi cell and weighted accordingly. Uniformly weighted input was applied, while a subset of Golgi cells were randomly inactivated to characterize the inhibitory effect caused by point coupling within the inhibitory nucleus; subsequently, M3 Golgi cells were randomly sampled from each granulocyte and weighted accordingly. Each inhibitory input is uniformly weighted and then used to regulate the excitatory input inheritance process of moss fibers, as shown in the following formula:

[0076]

[0077] GoCk (t)=0k=1,2...,or N2

[0078]

[0079] Where: N1 represents the total number of granulocytes; N2 represents the total number of Golgi cells; GoC k (t) represents the firing rate of the kth Golgi cell in the intermediate layer; θ j (t) represents the feedback inhibition experienced by the j-th granule cell; λ2 represents the feedback inhibition coupling strength;

[0080] S5: Purkinje cell layer completes task output. At each time point, the granule cell and its corresponding connection weight are multiplied sequentially, and the output results from all time windows are combined.

[0081] PC m (t)=∑w mj GC j (t)j=1,2,...,M3

[0082] Where: N3 represents the total number of Purkinje cells; W mj S6: Calculate the loss value and perform backpropagation. In a specific embodiment of the present invention, for handwritten digit speech transcription on the TI-46spoken wordcorpus speech dataset, the mean squared error (MSE) is selected as the loss function, as shown in the following formula:

[0083]

[0084] Where P(t) is the vector composed of the outputs of 3 Purkinje neurons, and T(t) is the expected 3D spatial trajectory vector.

[0085] According to the chain rule, the partial derivative of the loss function with respect to the firing frequency of Purkinje neurons is as follows:

[0086] Err(t) = P(t) - T(t)

[0087] Therefore, the changes in the connection weights between granulocytes and Purkinje cells can be calculated:

[0088] ΔW=W-(η*Err(t)*GC(t)′)

[0089] S7: Test network accuracy: If the cerebellar network accuracy is stable, stop network training. Once the network accuracy is stable or the maximum number of training iterations is reached, stop training and save the model at its optimal accuracy, such as... Figure 5 As shown.

[0090] Although the present invention has been described above, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many modifications under the guidance of the present invention without departing from the spirit of the present invention, and these modifications are all within the protection scope of the present invention.

Claims

1. A cerebellar dynamic sequence motion continuous learning control network, the control network comprising an input layer, an intermediate layer and an output layer; characterized in that, The intermediate layer is composed of a feedforward inhibition module and a feedback inhibition module; wherein: The input layer generates mossy fiber signals with dynamic timing; the dynamic timing mossy fiber signals are input into the network in time sequence as activated mossy fibers; The intermediate layer inputs the mossy fiber signals sparsely sampled by granule cells, while the inhibition module dynamically modulates the activity level and encoding mode of the granule cells; wherein: The feedforward inhibition module averages the firing rate of the mossy fiber signals to obtain the activity attribute of the Golgi cells inherited by the granule cells; ； ； wherein: is a piecewise linear activation function ; ; represents the number of moss fibers per cell sample of each granule; represents the total number of granule cells; The feedback inhibition module sparsely samples the granule cells to obtain the activity attribute of the Golgi cells inherited by the granule cells; Each Golgi cell was randomly sampled by the following equation one granule cell and weighted by uniformly weighted: ； wherein: represents the number of granulocytes per Golgi cell sample; represents the total number of Golgi cells; Each granule cell is randomly sampled by the equation one Golgi cell and weighted by uniformly weighted: ； ； wherein: represents feedback inhibition experienced by granulosa cells, represents the strength of the feedback inhibition coupling; The output layer processes and outputs the control cerebellar Purkinje cell output signal between the granule cell activity and the corresponding Purkinje cell at each time point according to the following formula: ； wherein: represents the total number of Purkinje cells, represents the firing rate of the th granule cell in the middle layer, represents the connection weight between the th granule cell and the th Purkinje cell.

2. The cerebellar model of a dynamic sequence learning control network according to claim 1, wherein, The process of the intermediate layer inputting the mossy fiber signals sparsely sampled by the granule cells, while the inhibition module dynamically modulating the activity level and encoding mode of the granule cells, comprises: Each granule cell is randomly sampled by the following equation The moss fiber signal is summed and weighted by The granule cell firing rate is obtained by uniform weighting ； wherein: is a piecewise linear activation function ; ; represents the number of moss fibers per particle cell sample; represents the total number of particle cells.

3. The method for controlling the cerebellar dynamics sequence motor learning continuity using the network according to claim 1, characterized in that, The process comprises the following steps: S1, building a cerebellar dynamic sequence motion control network based on firing rate; S2, intermediate layer granule cell sparse sampling: input into the mossy fiber of the input layer in time sequence, and the granule cells sparsely and uniformly receive the mossy fiber input; ； wherein, is a piecewise linear activation function ; denotes the total number of granule cells; denotes the number of mossy fiber synapses connected to each granule cell ； denotes the input layer mossy fiber synapse firing rate; S3, feedforward inhibition module granule cell input: Golgi cells average the excitatory input of the mossy fiber to form feedforward inhibition, and multiply by a scaling parameter to dynamically adjust the size and corresponding activation level of the granule cell input; that is: ； ； wherein: feed-forward inhibitory coupling strength; S4, feedback inhibition module granule cell input: Golgi cells sparsely sample the granule cells and sparsely act on the granule cells to form feedback inhibition, and multiply by a scaling parameter to dynamically adjust the size and corresponding activation level of the granule cell input; that is: ； ； ； ； wherein: represents the total number of granule cells; represents the total number of Golgi cells; represents the discharge rate of the intermediate layer Golgi cell; represents feedback inhibition experienced by the granule cell; represents the strength of the feedback inhibition coupling; S5, output layer Purkinje cell task output: multiply the connection weight between the granule cell activity after the inhibition module and the Purkinje cell to finally obtain the required output; that is: ； in: ; Indicates the first The granulocytes and the first Connection weights between Purkinje cells; S6, calculate loss value and gradient descent: count all the outputs of the output layer Purkinje cell at each time point, comprehensively calculate, and compare with the true target value to calculate the loss value; that is: ； wherein, is a vector of 3 Purkinje neuron outputs, is an expected 3-dimensional spatial trajectory vector; Update the connection weight of the granule cell and the Purkinje cell using the gradient descent learning algorithm; that is: According to the chain rule, the partial derivative formula of the loss function with respect to the Purkinje neuron firing frequency is as follows: ； ； S7, test network accuracy: if the network accuracy is stable or the number of training times reaches the maximum number, stop the network training.

Images