Model of brain-like expression and calculation unit
By simulating the information processing mechanism of human sensory organs, brain-like expression and calculation unit models are adopted, including brain-like units, compression-encoding bearing layer, sparse expression layer and self-organized mapping expression layer, solving the high requirements of existing brain-like models and The problem of poor interpretability achieves biorationality and interpretability, and is suitable for low-cost training of multimodal data.
Patent Information
- Application Number
- CN202411132430.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-11-19
AI Technical Summary
The existing brain-like model based on Transformer has shortcomings in terms of high data requirements and poor interpretability, especially in natural language processing tasks, which require a large amount of data and computing resources, and the internal structure of the model is complex and difficult to explain.
A model of brain-like expression and computing unit is proposed to simulate the information processing mechanism of human sensory organs, including brain-like units, compression-encoding bearing layer, sparse expression layer and self-organized mapping expression layer, using Elman recursive neural network and rewinding connection network. Build a computing architecture of biorational rationality and interpretability.
It achieves low data requirements and training energy consumption, provides biorationality and interpretability, is suitable for multimodal data, and reduces training complexity, and is suitable for large-scale brain-like expression and computing models.
Smart Images

Figure CN118982053A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to brain-like intelligent computing, and in particular to a model of brain-like expression and computing units. Background Art
[0002] Most existing brain-inspired models are based on Transformer, a deep learning architecture that is mainly used to process sequence data, such as natural language processing and image recognition. The Transformer model architecture includes self-attention and multi-head attention mechanisms, which enable the model to capture long-distance dependencies in the input sequence.
[0003] The following is a typical example of the application of the Transformer model architecture: Input sequence: First, the input sequence (such as text or image) is converted into embedding vectors. These embedding vectors will be used as input to the model.
[0004] Self-attention mechanism: The self-attention mechanism allows the model to assign different weights to each element in the input sequence in order to better capture the dependencies in the sequence. The self-attention mechanism calculates the similarity between each element in the input sequence and other elements and generates a weighted average for each element.
[0005] Multi-head attention mechanism: The multi-head attention mechanism applies self-attention to multiple representation spaces (called "heads") of the input sequence. Each head focuses on a different part of the input sequence, and their outputs are concatenated to obtain a richer representation.
[0006] Feedforward Neural Network: After the self-attention mechanism and the multi-head attention mechanism, the model further processes the representation of each element through a feedforward neural network. The feedforward neural network usually includes two linear transformations and a ReLU activation function.
[0007] Residual Connection Layer and Normalization: After the self-attention mechanism, multi-head attention mechanism, and feedforward neural network, the model uses the residual connection layer and normalization technology to improve the training stability and performance of the model.
[0008] Output: Finally, the output of the model can be used for various tasks such as text classification, machine translation, sentiment analysis, etc.
[0009] Transformer-based brain-like modeling has the following disadvantages: ① High data requirements. Transformer-type models usually require a large amount of data and computing resources for pre-training and fine-tuning. Especially in natural language processing tasks in some niche areas, more professional data and domain knowledge are often required to improve the generalization ability of the model; ② Poor interpretability. Since Transformer-type models use a self-attention mechanism to achieve encoding and representation learning of input sequences, the internal structure of the model is relatively complex, difficult to explain and understand, and lacks biological rationality and interpretability. Summary of the invention
[0010] In view of the above problems, this application aims to propose a model of brain-like expression and computing unit, which simulates the information processing mechanism of human sensory organs and has biological rationality and explainability.
[0011] The model of the brain-like expression and computing unit of the present application includes: a brain-like unit, a compression coding receiving layer; The brain-like unit includes: input layer, compression expression layer, sparse expression layer and self-organizing map expression layer; The compression coding receiving layer is used to store the compressed expression layer outputs of all adjacent brain-like units; The input layer is used to receive raw data or information; The compression expression layer is used to receive data from the input layer, the output of the compression expression layer of the brain-like unit at the previous moment, and the output of the compression expression layer of other neighboring brain-like units at the previous moment; The compression coding receiving layer is used to store the output of the compression expression layer; The sparse expression layer is used to receive the data of the compressed expression layer of the brain unit of this type for sparse coding, and the sparsely coded data is stretched into a sparse expression space; The self-organizing map expression layer is composed of a self-organizing map network, and 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 sparse expression layer into the self-organizing map expression space, thereby embedding the sparse expression space into the self-organizing expression space.
[0012] Preferably, the compression expression layer is an Elman recursive neural network.
[0013] Preferably, the compressed expression layers of neighboring brain-like units are connected to each other to form a spatiotemporal structure of the local context.
[0014] Preferably, the compression coding receiving layer is used for splicing and fusion between brain-like units and system integration, so that the compression expression layer can store and utilize information from past moments, thereby realizing the mapping of dynamic systems.
[0015] Preferably, the compression coding receiving layer receives the context-related expression coding data of the temporal state of the brain-like unit and the spatial state of other neighboring brain-like units to achieve learning and memory of similar or invariant patterns in time and space.
[0016] Preferably, the self-organizing expression layers corresponding to different brain-like units form a global communication network of the brain through a reentrant connection network.
[0017] Preferably, models of different brain-like expressions and computational units are distinguished in the form of tokens or in a splicing and fusion order.
[0018] Preferably, the expression layers of the models of different brain-like expression and computing units are projected to each other using a neural network.
[0019] Preferably, the raw data or information received by the input layer is analog data or a pulse sequence.
[0020] Preferably, when the raw data or information received by the input layer is analog data, it is pulsed in the self-organizing map expression layer.
[0021] The model of the brain-like expression and computing unit of the present application simulates the information processing mechanism of the human sensory organs based on the biological mechanism of brain sensory encoding, constructs a brain-like expression and computing unit architecture, and increases its biological rationality and explainability. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 It is a schematic diagram of the structure of the model of the brain-like expression and calculation unit of the present invention; Figure 2 A schematic diagram of the Elman recursive neural network structure of the compressed expression layer of the model of the brain-like expression and computing unit of the present invention; Figure 3 It is a schematic diagram of the splicing and fusion of the compression coding receiving layer between the units of the model of the brain-like expression and computing unit of the present invention; Figure 4 A schematic diagram of dynamic adjustment of reentrant connections between self-organizing expression layers between units of the model of the brain-like expression and computational unit of the present invention; Figure 5 A schematic diagram of the whole-brain full decoding encoding formed by the reentrant connections of functional columns in multiple cortical brain areas. DETAILED DESCRIPTION
[0023] The model of the brain-like expression and computing unit of this application, such as Figure 1As shown in the figure, it consists of two parts, namely brain-like units and compression coding succession layer. The brain-like unit includes input layer, compression expression layer, sparse expression layer and self-organizing map expression layer. The compression coding succession layer stores the compression expression layer output of all adjacent brain-like units. The brain-like unit simulates the information processing mode of the human brain sensory system.
[0024] Input layer: used to receive raw data or information.
[0025] Compressed expression layer: Elman recurrent neural network, such as Figure 2 It receives data from the input layer, the output of the compression expression layer of the unit at the previous moment, and the output of the compression expression layer of other adjacent brain-like units at the previous moment; the Elman recursive neural network compresses and encodes the data from the input layer; the Elman recursive neural network combines the output of the compression expression layer of the unit at the previous moment and the output of the compression expression layer of other adjacent brain-like units at this moment; the output of the compression expression layer is stored in the compression encoding receiving layer. The compression expression layers of neighboring units are connected to each other to form the spatiotemporal structure of the local context.
[0026] Compression coding receiving layer: stores the output of the compression expression layer, which is used for splicing and fusion between units and system integration, such as Figure 3 By introducing the continuation layer, the Elman recurrent neural network can store and utilize information from the past, thereby realizing the mapping of dynamic systems. The continuation layer here has a system integration and splicing fusion function very similar to that of the thalamus. It expresses the encoded data in a mosaic manner, such as a hexagonal mosaic, by contextually associating the temporal state of the continuation unit itself and the spatial state of the neighboring units, thus realizing the learning and memory of similar or invariant patterns in time and space.
[0027] Sparse expression layer: used for sparse coding, accepts data from the compressed expression layer, and then is independently and sparsely projected to the sparse expression layer.
[0028] Self-organizing map expression layer: It is composed of a self-organizing map network. The sparse expression layer is projected onto the self-organizing map expression layer, and the sparse expression space is embedded into the self-organizing expression space. The self-organizing expression layers corresponding to different brain-like units form a global brain communication network through a reentrant connection network, such as Figure 4 and Figure 5 shown.
[0029] Models of different brain-like expressions and computing units can be distinguished in the form of tokens or in the order of splicing and fusion; the expression layers of models of different brain-like expressions and computing units can be connected to each other by projecting them using commonly used neural networks.
[0030] The input data of the present invention can be analog data or a pulse sequence, and different brain-like units can receive a variety of different modal data inputs. If it is analog data, it needs to be pulsed by the self-organizing map expression layer after the sparse expression layer.
[0031] Solution 1: The input data is simulated data, which can be image data Assume that there are N brain-like units, and the i-th brain-like unit has ( arrive ) neighboring units, the input data is , the compressed projection matrix from the input layer to the compressed expression layer is , the projection matrix from the compression coding receiving layer to the compression expression layer is , the projection matrix from the compressed expression layer to the sparse expression layer is , the projection matrix from the sparse expression layer to the self-organizing map expression layer is Solution 1: Input data is simulated data, such as image data 1. Model initialization: , , , , the compressed expression layer output of the i-th brain-like unit in the compression coding receiving layer Assume the time step is T, starting from t=1: 2. Input Data After compression and projection, it reaches the compressed expression layer and is projected to obtain (t) 3. Calculate the compressed expression layer output: ,Will Save to 4. Calculate the sparse expression layer output: 5. Pulse: Transformed into a Poisson sequence 6. Calculate the output of the self-organizing map expression layer: 7. t+1 8. Go to 2 until t=T Solution 2: Input data is a pulse sequence of 0 and 1 1. Model initialization: , , , , the compressed expression layer output of the i-th brain-like unit in the compression coding receiving layer , let the time step be T, starting from t=1: 2. Input Data (t) After compression and projection, it reaches the compressed expression layer and is projected to obtain 3. Calculate the compressed expression layer output: ,Will Save to 4. Calculate the sparse expression layer output: 5. Calculate the output of the self-organizing map expression layer: 6. t+1 7. Go to 2 until t=T.
[0032] Taking vision as an example, the retina, as the receptor of the visual nervous system, is responsible for compressing and encoding external visual stimuli (corresponding to the compressed expression layer) and generating a data stream of neural pulse signals. These data streams are transmitted by the ganglion cell fibers of the retina and eventually sent to the dorsal thalamus and the lateral geniculate nuclei of the basal ganglia, forming the receptive field of the lateral geniculate nucleus in a highly ordered system separation and functional integration manner. In the lateral geniculate nucleus, the data streams from the adjacent retinal units and the visual cortex feedback are spliced and fused (corresponding to the compressed encoding receiving layer), making the visual information in the data stream more generalized and more potential. The data stream after the relay processing of the lateral geniculate nerve fibers is projected to the visual cortex, forming the projection field of the visual cortex in the form of independent sparse coding (sparse expression layer) and self-organizing mapping (corresponding to the self-organizing mapping expression layer). The lateral inhibition and reentrant connection mechanism between brain regions in the visual cortex brain area enable the brain to form a spatiotemporal activity brain topography and its dynamic connection network of synchronous or coherent oscillations within the brain region and between brain regions under the stimulation of the natural environment.
[0033] The model of the brain-like expression and computing unit of the present application simulates the information processing mechanism of human sensory organs and constructs a brain-like expression and computing unit architecture based on the biological mechanism of brain sensory encoding. It has strong biological rationality and interpretability, and the model can be trained using unsupervised training methods. The input data can use multimodal data, which has low requirements on data and does not require additional labeling. The model input can be either simulated data or a pulse sequence, which reduces the training energy consumption to a certain extent and lays the foundation for large-scale brain-like expression and computing models.
Claims
1. A model of a brain-like expression and computational unit, comprising: Brain-like unit, compression coding receiving layer; The brain-like unit includes: input layer, compression expression layer, sparse expression layer and self-organizing map expression layer; The compression coding receiving layer is used to store the compressed expression layer outputs of all adjacent brain-like units; The input layer is used to receive raw data or information; The compression expression layer is used to receive data from the input layer, the output of the compression expression layer of the brain-like unit at the previous moment, and the output of the compression expression layer of other neighboring brain-like units at the previous moment; The compression coding receiving layer is used to store the output of the compression expression layer; The sparse expression layer is used to receive the data of the compressed expression layer of the brain unit of this type for sparse coding, and the sparsely coded data is stretched into a sparse expression space; The self-organizing map expression layer is composed of a self-organizing map network, and 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 sparse expression layer into the self-organizing map expression space, thereby embedding the sparse expression space into the self-organizing expression space.
2. The model of the brain-like expression and computing unit according to claim 1, characterized in that: The compression expression layer is an Elman recurrent neural network.
3. The model of the brain-like expression and computing unit according to claim 2, characterized in that: The compressed representation layers of neighboring brain-like units are connected to each other to form the spatiotemporal structure of the local context.
4. The model of the brain-like expression and computing unit according to claim 2, characterized in that: The compression coding receiving layer is used for splicing, fusion and system integration between brain-like units, so that the compression expression layer can store and utilize information from past moments, thereby realizing the mapping of dynamic systems.
5. The model of the brain-like expression and computing unit according to claim 4, characterized in that: The compression coding receiving layer receives the context-related expression coding data of the temporal state of the brain-like unit and the spatial state of other neighboring brain-like units to achieve learning and memory of similar or invariant patterns in time and space.
6. The model of the brain-like expression and computing unit according to claim 1, characterized in that: The self-organizing expression layers corresponding to different brain-like units form a global communication network of the brain through a reciprocal connection network.
7. The model of the brain-like expression and computing unit according to claim 1, characterized in that: Models of different brain-like expressions and computational units are distinguished by tokens or splicing and fusion orders.
8. The model of the brain-like expression and computing unit according to claim 1, characterized in that: The expression layers of the models of different brain-like expression and computing units are projected onto each other using neural networks.
9. The model of the brain-like expression and computing unit according to claim 1, characterized in that: The raw data or information received by the input layer is analog data or pulse train.
10. The model of the brain-like expression and computing unit according to claim 1, characterized in that: When the raw data or information received by the input layer is analog data, it is pulsed in the self-organizing map expression layer.
Citation Information
Patent Citations
Computer system based on brain-like coprocessor
CN104809501A
Brain-like computing element architecture model and brain-like computing device comprising same
CN118057472A
Pattern generating method using self-organizing map, its program and device
JP2008071255A
Controller optimization via reinforcement learning on asset avatar
US20200370423A1
