Neuromorphic IC Spiking Neural Network Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning applications face challenges due to the need for extensive annotated training data, subject matter expertise, efficient real-time implementation on spiking neural networks (SNN), porting of SNN weights to neuromorphic computers/ICs, and statistical reliability testing, which are costly and require significant resources.
Innovation Solution
A novel method that generates annotated training data, converts deep neural networks (DNNs) to SNNs, implements them on neuromorphic ICs, and ensures reliable performance through high-fidelity sensor models and statistical testing, enabling efficient deep learning AI solutions for advanced sensor signal processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning networks are implemented using conventional architectures, then computational accuracy is improved, but real-time processing capability deteriorates
Solution Approach 1:
The patent replaces conventional von Neumann architecture with spiking neural network architecture that mimics biological neural processing. This substitution enables event-driven asynchronous computation where neurons communicate via spikes only when necessary, achieving both high accuracy and real-time processing efficiency simultaneously
2Reliability
If extensive annotated training data is used to ensure DNN convergence, then model performance is improved, but data preparation cost and time deteriorate
Solution Approach 1:
The patent uses synthetic data generation through physics-based sensor models to create training datasets. These simulated datasets copy the essential characteristics of real sensor data without requiring actual annotated real-world samples, dramatically reducing data preparation time while maintaining model performance
Solution Approach 2:
The patent performs preliminary data generation and model training using synthetic datasets before deployment. This preliminary action with simulated data prepares the model in advance, eliminating the need for extensive real-world data collection and annotation during deployment phases
3Use of energy by moving object
If DNN is converted to SNN for neuromorphic implementation, then energy efficiency is improved, but conversion complexity deteriorates
Solution Approach 1:
The patent transforms DNN to SNN by changing key parameters: converting continuous activation functions to discrete spike events, changing weight representations to spike timing encodings, and modifying the temporal dynamics from synchronous to asynchronous. These parameter changes enable energy-efficient neuromorphic implementation while systematic conversion protocols manage the complexity
Data Source
AI summary
A novel system and method are described that allows for implementation of compact and efficient deep learning AI solutions to advanced sensor signal processing functions. The process includes the following stages: (1) A method for generating requisite annotated training data in sufficient quantity to ensure convergence of a deep learning neural network (DNN); (2) A method for implementing the resulting DNN onto a Spiking Neural Network (SNN) architecture amenable to efficient neuromorphic integrated circuit (IC) architectures; (3) A method for implementing the solution onto a neuromorphic IC; and (4) A statistical method for ensuring reliable performance.