Firing Neural Network Modeling Across Scales and Topologies
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing brain-like firing neural network computation frameworks lack a unified and flexible modeling method, fail to support flexible network topological structures, do not integrate sufficient built-in functions, are not compatible with various chip architectures, and struggle with efficient medium or large-scale deployment on general-purpose computers or embedded devices.
Innovation Solution
A flexible and efficient firing neural network computing system with a multi-level tree structure that supports full-scale modeling, integrates traditional neural networks, and includes automatic synapse and neuron pruning, enabling compatibility with CPU, ARM, GPU, DSP, and other processors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing brain-like firing neural network computation frameworks use fixed modeling methods, then implementation is simpler, but modeling flexibility and support for multiple scales is reduced
Solution Approach 1:
The patent implements a universal computing node that can function as different types of neurons (spiking neurons, rate neurons, etc.) and support multiple computing modes (event-driven, tensor-driven) through configurable parameters. This allows a single framework to handle various neural network models and scales without requiring separate implementations for each type, thereby improving adaptability while managing complexity through unification.
Solution Approach 2:
The patent employs a hierarchical nested structure where computing nodes are organized in multiple levels: individual neurons are nested within populations, populations are nested within brain regions, and brain regions are nested within the complete neural network. This nested organization allows the system to model multiple scales simultaneously (from single neuron to whole brain) while maintaining a unified framework, resolving the contradiction between modeling flexibility and system complexity.
2Adaptability or versatility
If existing frameworks support limited network topologies, then system complexity is lower, but adaptability to different brain structures is reduced
Solution Approach 1:
The patent implements a universal connection mechanism that can represent various network topologies (feedforward, feedback, lateral connections, recurrent connections) through a unified connection definition interface. The computing nodes can be connected in arbitrary patterns by specifying source and target nodes, allowing support for diverse brain structures without requiring separate handling for each topology type, thus improving adaptability while controlling complexity through unification.
3Ease of operation
If existing frameworks lack automatic pruning and genesis functions, then system complexity is lower, but ease of operation is reduced
Solution Approach 1:
The patent implements automatic pruning and genesis functions that enable the neural network to self-optimize its structure during training and operation. The system automatically prunes redundant connections based on connection strength thresholds and generates new connections based on learning requirements, eliminating the need for manual intervention. This self-service capability improves ease of operation while the automated nature of these functions manages complexity by using algorithmic rules rather than manual configuration.
4Productivity
If existing frameworks are not optimized for hardware compatibility, then portability is higher, but computing efficiency is reduced
Solution Approach 1:
The patent implements hardware-specific parameter optimization where the same neural network model can be configured with different computational parameters depending on the target hardware platform. For example, precision parameters (float32, float16, int8), batch size, and parallelization settings can be adjusted based on the capabilities of CPU, GPU, or other processors. This allows the system to achieve high computing efficiency on different hardware platforms while maintaining the same underlying model architecture, thus improving productivity without sacrificing hardware compatibility.
Data Source
AI summary
A firing neural network computing system and method for brain-like intelligence and cognitive computing. The system includes a model description module, a parameter database, a configuration description module, a configuration manager, a rule manager, a data manager, a network builder, a network manager, an operation manager, a scheduler, a log manager, an operation monitoring module and a graphical display module. The system provides the function of automatically executing synapse and neuron pruning and genesis according to certain conditions and rules, and provides a variety of flexible trigger conditions for starting up pruning and genesis processes as well as rules for executing the processes, which eliminates the burden of neural network developers needing to write synapse and neuron pruning and genesis programs by themselves, thereby effectively solving several problems in existing brain-like firing neural network computing frameworks.