Probabilistic Spike Neural Network Circuit for Low-Power Fast Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
General spike neural networks perform complex multiplication and accumulation operations, leading to increased power consumption and reduced operation speed.
Innovation Solution
Incorporation of a probabilistic operator in the spike neural network circuit that simplifies operations by using a conversion table to convert fire probabilities into threshold values and utilizes a random number generator and comparator to generate output spike signals based on comparison with these threshold values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If general spike neural network performs multiplication operation of weight and accumulation operation of charge, then the neural network can process information, but the operation procedure becomes complicated and power consumption increases
Solution Approach 1:
The patent transforms the conventional multiplication and accumulation operations into a probabilistic operation where the weight value determines a fire probability. The conversion table converts weight values to threshold values, and a random number generator produces random values. The comparison between random value and threshold determines whether the neuron fires, replacing complex arithmetic with simpler probabilistic sampling that consumes less power and has fewer operational steps
Solution Approach 2:
The patent replaces the mechanical arithmetic operations (multiplication and accumulation) with a probabilistic mechanism involving random number generation and threshold comparison. This substitution eliminates the need for complex computational hardware while achieving the same neural network functionality through stochastic processes
2Speed
If general spike neural network performs multiplication and accumulation operations, then the neural network can generate output spikes, but the operation speed is reduced
Solution Approach 1:
The patent changes the operational parameters from deterministic arithmetic operations to probabilistic sampling. By converting weight values to threshold values through a lookup table and using random number generation, the system achieves faster operation speeds because probabilistic sampling requires fewer computational steps compared to traditional multiplication and accumulation operations
3Ease of operation
If spike neural network uses probabilistic operator with random number generation, then operation procedure is simplified and power consumption is reduced, but additional components are required
Solution Approach 1:
The patent merges the conversion table and random number generator into an integrated probabilistic operator module. The conversion table is implemented as a lookup structure within the circuit, and the random number generator is combined with the threshold comparison logic, creating a unified component that simplifies the overall circuit architecture while maintaining operational simplicity
Data Source
AI summary
Disclosed is a spike neural network circuit including an axon circuit that generates a first input spike signal, a conversion table that converts a first fire probability of a first neuron corresponding to the first input spike signal into a first threshold value, and a probabilistic operator. The probabilistic operator includes a random number generator that generates a random number value based on an event that the first input spike signal is at a first logic level, a random number comparator that generates a first comparison signal by comparing the first threshold value with the random number value, and a spike generator that generates an output spike signal corresponding to the first neuron based on an event that the first comparison signal is at the first logic level.