Tabu Search Mapping for Spiking Neural Network On-Chip Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing mapping optimization methods for spiking neural networks on-chip, such as those using greedy algorithms, fail to consider the global connections and dynamic spike frequencies within complex neural networks, leading to inefficiencies in computing and energy consumption.
Innovation Solution
A neural network on-chip mapping method based on a tabu search algorithm, which involves acquiring topological structures and parameter information of the spiking neural network, converting it into a directed acyclic graph, and using a nearest neighbor algorithm and integer programming to optimize core mapping and reduce communication costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a greedy algorithm is used for mapping optimization, then the mapping process is simple and fast, but it fails to consider global connections and dynamic spike frequencies, leading to high communication costs and wasted computing efficiency
Solution Approach 1:
The patent employs the Tabu search algorithm to dynamically adjust the mapping configuration based on real-time communication costs and spike frequencies. The algorithm iteratively explores different mapping solutions, evaluating communication overhead and computational efficiency, then transitions to superior configurations while avoiding cycles through tabu list mechanisms. This dynamic optimization process enables the system to adapt to varying neural network workloads and achieve optimal resource allocation.
Solution Approach 2:
The patent transforms the mapping optimization problem by changing key parameters: instead of using static greedy heuristics, the system uses Tabu search with iterative parameter adjustments. The algorithm modifies mapping assignments based on evaluated communication costs, spike firing frequencies, and connection patterns, continuously improving the solution space to find near-optimal mappings that balance communication efficiency and computational performance.
2Device complexity
If a greedy algorithm is used for mapping, then the computation is simple, but it does not fully consider global connections and irregular sparse connections in complex neural networks
Solution Approach 1:
The Tabu search algorithm provides dynamic exploration of the solution space, allowing the mapping to adapt to the complex global connection structure of spiking neural networks. The algorithm iteratively evaluates different mapping configurations considering both local and global connection patterns, ensuring high mapping quality that accounts for irregular sparse connections and global dependency relationships.
Solution Approach 2:
The patent implements feedback mechanisms where the mapping quality is continuously evaluated based on communication costs, spike frequencies, and connection patterns. The Tabu search algorithm uses this feedback to guide the search process, accepting improvements and rejecting deteriorations, thereby ensuring reliable mapping quality that fully considers global connections and network topology.
3Loss of energy
If segmentation and redistribution are performed to reduce input mapping efficiency ratio, then redundant bandwidth is reduced, but the mapping process becomes more complex and requires additional optimization steps
Solution Approach 1:
The patent uses Tabu search to dynamically optimize the segmentation and redistribution process. The algorithm iteratively adjusts mapping assignments to minimize redundant bandwidth while managing the complexity of the mapping process. By evaluating communication patterns and spike frequencies, the system finds optimal segmentation strategies that reduce bandwidth overhead without requiring overly complex manual intervention.
Data Source
AI summary
A neural network on-chip mapping method and apparatus based on a tabu search algorithm are provided. The method includes: constructing a tabu search table and using a heuristic-based iterative search process to select local computing cores of a network-on-chip as candidates, establishing an integer programming model and solving an optimal solution, continuously reducing an objective cost function of a mapping solution by loop iteration, and finally obtaining an approximately optimal deployment scheme.