Spiking Neural Network Feature Extraction via Dynamic Reconfiguration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current artificial neural networks, including spiking neural networks, face challenges in replicating the functionality of biological neural networks, particularly in scaling to large networks, making rapid inferences from diverse input data, and allowing for user interventions to update trained models effectively.
Innovation Solution
A neuromorphic system with a spike converter, reconfigurable neuron fabric, and processor is developed, enabling a spiking neural network to learn and perform unsupervised, semi-supervised, and supervised feature extraction by modifying synaptic weights based on input and user input, with asynchronous spiking and synchronous winning neuron selection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If traditional artificial neural networks are used to replicate biological neural network functionality, then computational capability is improved, but power consumption and device complexity increase significantly
Solution Approach 1:
The patent replaces traditional von Neumann architecture with a neuromorphic system that uses event-driven spiking neurons and synapses. Information is transmitted via discrete spike events rather than continuous data flow, enabling asynchronous operation that consumes power only during computation events, thus dramatically reducing overall power consumption while maintaining computational capability.
Solution Approach 2:
The system uses periodic spiking activity where neurons fire discrete action potentials at specific moments rather than continuously processing data. This event-driven periodic action allows the network to remain in low-power states between spikes, achieving efficient computation with significantly reduced energy consumption compared to traditional continuous processing architectures.
2Extent of automation
If spiking neural networks are scaled to very large networks, then computational power is improved, but hardware complexity and difficulty of implementation increase
Solution Approach 1:
The patent divides the large-scale neural network into modular functional units including spike converters for different sensor types (camera, microphone, DVS), multiple layers of spiking neurons, and synaptic layers. Each module is independently configurable and can be scaled by adding or removing modules, making large-network implementation manageable through systematic decomposition.
Solution Approach 2:
The system employs universal spiking neuron circuits and synapse circuits that can handle multiple types of input data (visual, auditory, depth) through configurable spike converters. The same core neural processing architecture processes all sensor inputs, reducing hardware complexity by avoiding specialized circuits for each data type while maintaining versatility.
3Speed
If traditional neural networks make rapid inferences from diverse input data, then processing speed is improved, but adaptability to update trained models decreases
Solution Approach 1:
The patent implements dynamic reconfigurability where the neural network architecture can be modified during operation. Users can add new spiking neurons, adjust synaptic weights, and reconfigure connections between layers based on incoming data patterns. This dynamic adaptation allows the system to maintain rapid inference speeds while continuously learning from new data sources and updating its computational model.
Data Source
AI summary
Disclosed herein are system, method, and computer program embodiments for an improved spiking neural network (SNN) configured to learn and perform unsupervised, semi-supervised, and supervised extraction of features from an input dataset. An embodiment operates by receiving a modification request to modify a base neural network, having N layers and a plurality of spiking neurons, trained using a primary training dataset. The base neural network is modified to include supplementary spiking neurons in the Nth or N + 1th layer of the base neural network. The embodiment includes receiving a secondary training dataset and determining membrane potential values of one or more supplementary spiking neurons in the Nth or Nth + 1 layer which learn features based on secondary training data set to select a supplementary/winning spiking neuron. The embodiment performs a learning function for the modified neural network based on the winning spiking neuron.