Event-Driven Spiking Neural Accelerator for Inhibitory Jitter
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing spiking neural network accelerators face inefficiencies due to high computation delay and power consumption, particularly in event-driven models, and the spiking jitter problem in inhibitory networks, leading to errors in computation results.
Innovation Solution
An approximate computation model is constructed based on spiking neuron models and coding modes, ignoring fine-grained time information to reduce computation dependency and spiking jitter by collecting signals in coarse-grained time slices, computing membrane voltage gain, and determining spiking firing frequency, supported by an event-driven accelerator with modules for spike input, control, state, and output processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional event-driven accelerator processes inhibitory spiking neural network, then computation amount is reduced, but spiking jitter problem occurs causing computation errors
Solution Approach 1:
The patent applies preliminary action by collecting all spiking signals into a time slice before processing, and by delaying the transmission time of spiking signals until the final state is determined. This ensures that neurons are processed in a consistent state, preventing spiking jitter and ensuring accurate computation results while maintaining event-driven efficiency
Solution Approach 2:
The patent uses feedback mechanisms through the bitmap data structure to track neuron states and determine whether to transmit spiking signals. The system continuously monitors membrane voltage states and uses this feedback to make informed decisions about signal transmission, ensuring accuracy while maintaining efficiency
2Ease of manufacture
If time-driven accelerator scans all neuron states at end of each time step, then implementation is simple, but redundant computations occur
Solution Approach 1:
The patent extracts only the necessary information (spiking signals) from the complete neuron state space and processes only those events. By taking out the essential spiking events and processing them through event-driven mechanisms, the system achieves both simplicity in implementation and efficiency in computation, avoiding redundant scans of all neuron states
Solution Approach 2:
The patent segments the computation process into discrete events based on spiking signals rather than continuous time steps. This segmentation allows the system to process only relevant computations when spiking events occur, maintaining implementation simplicity while dramatically reducing redundant computations
3Measurement precision
If spiking neuron membrane voltage updated multiple times in time domain, then computation accuracy is ensured, but computation amount increases
Solution Approach 1:
The patent merges multiple time-step computations into a single event-driven processing cycle. By collecting all spiking signals into a time slice and processing them together, the system achieves the same computational accuracy as multiple sequential updates but with significantly reduced total computation amount, as all necessary computations are performed in parallel within the merged time slice
Data Source
AI summary
A spiking neural network acceleration method includes constructing an approximate computation model according to a spiking neuron model and a spiking coding mode. The approximate computation model utilizes the characteristic that spiking frequency domain coding ignores the time semantics of a spiking sequence, compresses the distribution of spiking signals on a time step, and greatly reduces a spiking routing process and a subsequent neural computation process. The time-driven accelerator replans the computation process of the spiking neural network, sets a deduplication queue and a bitmap to solve the problem of spiking jitter, and realizes efficient support on the inhibitory spiking neural network.