RL-NAS Spiking Neural Network Generation for Neuromorphic Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating Spiking Neural Networks (SNNs) for neuromorphic devices face challenges such as complexity in neuron models, training and inference inefficiencies, model selection for high accuracy, computational inefficiencies, and hardware implementation incompatibilities, leading to sub-optimal performance and resource constraints.
Innovation Solution
A method and system utilizing Reinforcement Learning (RL) Neural Architecture Search (NAS) with constrained search spaces and Faster-NAS techniques to generate optimized SNNs, incorporating neuromorphic hardware constraints, optimizing NPU utilization, and converting DNN models to SNNs efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional handcrafting approach is used for SNN model, then model can be customized for specific tasks, but the process is time-consuming and requires subject matter expert dependency
Solution Approach 1:
The system enables automated self-service for SNN model creation through neural architecture search. The NAS algorithm automatically searches the architecture space, evaluates candidate models, and selects optimal architectures without requiring manual expert intervention, thereby eliminating SME dependency and reducing creation time
Solution Approach 2:
An automated neural architecture search system acts as an intermediary between the design requirements and the final SNN model. This intermediary automatically performs architecture exploration, evaluation, and selection, replacing the need for direct human expert involvement in the model creation process
2Extent of automation
If ANN2SNN conversion is used, then automation is improved, but conversion errors and information loss occur especially under low time steps
Solution Approach 1:
The system performs preliminary architecture search and optimization in the SNN domain before final model deployment. By searching the architecture space with SNN-specific constraints and evaluation metrics beforehand, the system identifies architectures that are inherently suitable for spiking computation, preventing conversion errors and information loss
Solution Approach 2:
The system changes the fundamental approach from converting ANNs to SNNs to directly searching for optimal SNN architectures. This parameter change in the generation methodology allows the model to be natively designed for spiking neurons, eliminating conversion-related accuracy degradation and information loss
3Extent of automation
If SNN search space exploration is performed, then automated model generation is achieved, but hardware constraints are not met or accuracy is lost
Solution Approach 1:
The system applies local quality by incorporating hardware-specific constraints into the architecture search process. Different candidate architectures are evaluated against specific hardware metrics such as NPU utilization, memory requirements, and power consumption, allowing the selection of models that are locally optimized for target hardware platforms
Solution Approach 2:
The system implements feedback mechanisms where candidate architectures are evaluated based on both accuracy metrics and hardware constraint satisfaction. The evaluation results feed back into the search process, guiding the exploration toward architectures that simultaneously meet accuracy requirements and hardware limitations, thereby resolving the contradiction between automation and reliability
4Measurement precision
If complex neuron models are used, then simulation accuracy is improved, but computational efficiency and training difficulty increase
Solution Approach 1:
The system changes the approach from manually selecting complex neuron models to automatically searching for optimal architecture configurations. The neural architecture search process evaluates different model complexities and selects architectures that achieve the desired simulation accuracy while maintaining computational efficiency, thereby resolving the trade-off between precision and productivity
Data Source
Figure 1A
Figure 1B
Figure 2A
AI summary
State of art techniques such as ANN2SNN yield sub-optimal performance for Spiking Neural Networks (SNNs), while with SNN search space Neural Architecture Search (NAS) based approach many times the target SNN hardware constraints might not be met or there can be accuracy loss. A method and system for rapid automated generation of optimized SNNs for neuromorphic devices is disclosed. A Reinforcement Learning (RL)-NAS technique is utilized to obtain Deep Neural Networks (DNNs) using raw training dataset in non-spike data format and set of SNN constraints in accordance with the neuromorphic hardware of target neuromorphic devices along with general target hardware constraints, wherein the set of constraints are imbibed into the NAS space in form of constrained graph. An optimal DNN is model searched in an SNN contained NAS space and trained on the raw data. An optimal SNN model is obtained from the trained DNN model via neuromorphic OEM converter.