A multi-axis multi-directional neural network topology based on a 3D coordinate system

Abstract

The application relates to the technical field of pulse neural networks, and discloses a pulse neural network topological structure, a pulse neural network, a face binding memory communication method and a system. The topological structure adopts a nine-axis extension structure of a three-dimensional coordinate system, each neuron is connected with 18 direction neighbors, and the axes can be extended to six axes, twelve axes and higher; the nine axes include three main axis directions and six face diagonal directions. The network supports multi-face parallel signal receiving, realizes multi-modal information synchronous coding propagation, and each face can be independently configured as an input / output face and bound to different signal sources. The communication method takes a specific face as a memory interaction interface, and neurons in the face form a continuous context space, supporting large-scale memory parallel writing. The system integrates the above structure and mechanism, improves information processing and memory interaction efficiency, and is suitable for multi-modal information processing scenes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence neural network technology, specifically to a spiking neural network topology based on a multi-axis three-dimensional lattice structure, a multi-faceted parallel input mechanism, a faceted bound memory communication method and system. Background Technology

[0002] While some existing brain-inspired models incorporate the concept of functional partitioning, they still exhibit fixed layer structures, failing to autonomously adjust their network morphology based on the characteristics of the input data. Furthermore, most spiking neural networks rely on pre-defined layer structures and connection paths, hindering data-driven self-organizing partitioning. This results in insufficient network flexibility, weak generalization ability, and difficulty in meeting the demands of complex tasks. Moreover, existing technologies lack spiking neural networks based on 3D lattice structures, without fixed layers, capable of autonomously forming functional partitions, thus failing to fully simulate the self-organizing and adaptive characteristics of the human brain. Summary of the Invention

[0003] The present invention aims to overcome the shortcomings of the prior art and provide a self-organizing brain-like spiking neural network based on a multi-axis 3D lattice, which realizes high spatial efficiency, multimodal parallel processing and surface-bound memory communication.

[0004] This invention discloses a 9-axis extended topology for a three-dimensional coordinate system, including 3 principal axis directions ±i, ±j, and ±k, and 6 face diagonal directions ±(i±j), ±(i±k), and ±(j±k), for a total of 18 neighbor connection directions. The number of axes can be flexibly expanded to 6 axes, 12 axes, and higher dimensions. The network abandons fixed layers and uses 3D lattice neurons as basic units, autonomously forming functional partitions based on the characteristics of the input signal.

[0005] The network supports multi-faceted parallel input and output, with each facet independently configurable as an input facet, output facet, or memory interface facet. Different faces are bound to different signal sources, enabling synchronous encoding and propagation of multimodal information. Employing a facet-bound memory communication method, a designated network facet serves as a memory interaction interface, with neurons within the facet forming a continuous context space, supporting large-scale parallel writing and reading of memory data.

[0006] This invention features a regular structure, stable connections, and efficient propagation. It can self-organize and adapt to tasks, significantly improving the flexibility and memory interaction capabilities of brain-like computing.

[0007] Specific implementation methods (1) 3D multi-axis lattice topology construction: neurons are distributed in three-dimensional lattice space, and each neuron establishes 18 fixed-direction neighbor connections according to the 9-axis rule; the number of axes can be expanded to 6 axes and 12 directions or 12 axes and 24 directions according to computing power and task. (2) Multi-face parallel input and output: external multimodal signals are injected in parallel from different faces of the network after pulse coding. Each face is configured with independent functions and does not interfere with each other, realizing synchronous processing of visual, text and other multimodal information. (3) Self-organizing functional partitioning: during network operation, stable functional areas are automatically formed according to the signal distribution and pulse propagation law, without the need for preset hierarchy or partitioning rules. (4) Face-bound memory communication: a specified face is set as a memory interface, and external memory is written into the lattice in batches through this face. The neuron context is continuously stored; the processing result is read by the output face, realizing high-speed, parallel, and low-latency memory interaction. (5) System integration: The system integrates multi-axis 3D topology, multi-faceted parallel input, self-organizing partitioning and face binding memory mechanism to form a complete brain-like spiking neural network system, which is suitable for multimodal perception, autonomous decision-making and continuous memory scenarios.

Claims

1. A spiking neural network topology, characterized in that: A multi-axis extended topology in a three-dimensional coordinate system is adopted, in which each neuron is connected to its neighbors in 18 directions; the number of axes can be expanded as needed, including 6 axes with 12 directions, 12 axes with 24 directions, and higher axis numbers.

2. The topology according to claim 1, characterized in that: The 18 directions consist of 9 axes, including 3 orthogonal principal axis directions (±i, ±j, ±k) and 6 face diagonal directions (±(i±j), ±(i±k), ±(j±k)).

3. A spiking neural network, characterized in that: It has multiple network surfaces, which can receive external signal inputs in parallel to realize the synchronous encoding and propagation of multimodal information.

4. The spiking neural network according to claim 3, characterized in that: Each network surface can be independently configured as an input or output surface, and different network surfaces can be bound to different signal sources or functions.

5. A surface-bound memory communication method, characterized in that: A specified network surface of the spiking neural network is bound as a memory interaction interface. External memory data is written into the neural network through this memory interaction interface, and the processing results of the neural network are read through the output surface.

6. The method according to claim 5, characterized in that: All neurons within the memory interaction interface plane constitute a continuous context space, supporting the parallel writing of large-capacity memory data.

7. A spiking neural network system, characterized in that: It includes the topology described in claims 1-2, the multi-faceted parallel input mechanism described in claims 3-4, and the face-bound memory communication method described in claims 5-6. 8.9 Axis 18 Direction Definition Table: