Machine Learning Circuit Neuron Cell Lookup Table Energy Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural network circuits face high energy inefficiency due to the necessity of multiply and accumulation calculations and frequent weight information storage and retrieval operations.
Innovation Solution
A machine learning processing circuit is designed with neuron cell circuits that include an input unit, an adder unit for accumulating input signals, and a storage unit that holds output results of non-linear functions, allowing for reduced memory readings and calculations, thereby enhancing energy efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If conventional neural network circuits perform multiply and accumulation calculations with weight information, then calculation accuracy is maintained, but energy consumption increases significantly
Solution Approach 1:
The patent extracts the multiplication operation from the neural network calculation and replaces it with a lookup table approach. Weight values are pre-stored in memory, and during inference, instead of performing multiply-accumulate operations, the system directly retrieves pre-computed values from the lookup table based on input signals, thereby eliminating complex multiplication while maintaining calculation accuracy.
Solution Approach 2:
The patent applies preliminary action by pre-computing and storing weight values in lookup tables before the actual neural network inference. During operation, the system only needs to perform simple table lookups and additions rather than complex multiplications, significantly reducing real-time computational energy consumption while preserving the neural network's functional accuracy.
2Use of energy by moving object
If conventional circuits perform frequent weight information storage and retrieval, then neural network functionality is maintained, but energy efficiency deteriorates
Solution Approach 1:
The patent merges the weight storage function with the computation function by integrating lookup tables directly into the neural network circuit architecture. Instead of separate weight memory and computation units, the weight values are embedded within the computation pathway, allowing simultaneous access to weight information during inference without requiring additional memory read/write operations.
3Measurement precision
If multiply and accumulation calculations are performed for each neuron, then accurate neural network output is achieved, but computational cost increases
Solution Approach 1:
The patent uses copying by creating lookup tables that contain pre-stored weight values and activation function results. Instead of performing actual multiplication operations during inference, the system copies relevant values from the lookup tables based on input signals, thereby maintaining calculation accuracy while dramatically improving computational efficiency through simplified data retrieval operations.
Data Source
AI summary
A machine learning processing circuit including a plurality of neuron cell circuit. Each of the plurality of neuron cell circuit includes an input unit that receives a plurality of input signals, an adder unit that adds the input signals received by the input unit, and a storage unit that holds output results of a non-linear function corresponding to input values, uses an output signal output by the adder unit, as an input value, and outputs an output result of the non-linear function corresponding to the input value.