Reconfigurable Spiking Neural Network for Low-Power Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional artificial neural networks, including spiking neural networks, face challenges such as high power consumption, inability to switch between convolution and fully connected operations, lack of reconfigurability, limited synapse weight sharing, and inefficiencies in processing and learning methods, which hinder their ability to efficiently perform unsupervised feature extraction and inference tasks.
Innovation Solution
A neuromorphic integrated circuit with a reconfigurable spiking neural network architecture that incorporates a spike converter, a reconfigurable neuron fabric, memory, and a processor, enabling unsupervised feature extraction through Spike Time Dependent Plasticity (STDP) learning and allowing for configuration changes via user-defined files, with features like ternary synapse weights, input and output spike buffers, and synapse weight sharing to enhance processing efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If conventional artificial neural networks are used, then they can perform basic neural network operations, but they consume high power and cannot efficiently perform unsupervised feature extraction
Solution Approach 1:
The patent replaces conventional von Neumann architecture with a neuromorphic architecture that mimics biological neural networks. The spiking neural network uses event-driven computation where neurons communicate via discrete spikes rather than continuous values, and synaptic weights are modified through STDP learning rules that operate locally and asynchronously. This substitution of computational paradigm reduces power consumption by several orders of magnitude while enabling efficient unsupervised feature extraction through biologically-inspired learning mechanisms.
Solution Approach 2:
The patent implements a reconfigurable neural network architecture where the computational graph can be dynamically modified at runtime. The system can switch between different network configurations (e.g., convolutional, fully connected, recurrent) and adjust synaptic connections based on learning needs. This dynamic reconfigurability allows the network to adapt its structure for different tasks, improving both energy efficiency and feature extraction performance compared to static conventional networks.
2Adaptability or versatility
If conventional neural networks are used, then they have fixed architecture, but they lack reconfigurability and cannot switch between convolution and fully connected operations
Solution Approach 1:
The patent designs a universal neuromorphic processing unit that can perform multiple neural network operations including convolution, fully connected layers, and recurrent operations within the same hardware structure. The reconfigurable synaptic array and programmable neuron circuits allow a single device to emulate different network architectures and operations, providing versatility without requiring separate dedicated hardware for each operation type.
Solution Approach 2:
The patent divides the neural network into modular components: input spike buffers, configurable synaptic layers, neuron processing units, and output buffers. Each module can be independently configured and reconnected through a programmable interconnect fabric. This segmentation allows flexible reconfiguration of the computational graph while maintaining manageable device complexity through standardized interfaces and modular design.
3Productivity
If conventional spiking neural networks are used, then they implement basic spiking behavior, but they have limited synapse weight sharing and inefficient learning methods
Solution Approach 1:
The patent implements Spike Time Dependent Plasticity (STDP) learning where synaptic weights are dynamically adjusted based on the relative timing of pre- and post-synaptic spikes. This local Hebbian learning rule provides efficient unsupervised learning without requiring backpropagation through the entire network. The feedback mechanism operates asynchronously and locally at each synapse, improving learning efficiency while enabling synapse weight sharing through common weight parameters that can be updated by multiple neurons.
Data Source
AI summary
Disclosed herein are system, method, and computer program product embodiments for an improved spiking neural network (SNN) configured to learn and perform unsupervised extraction of features from an input stream. An embodiment operates by receiving a set of spike bits corresponding to a set synapses associated with a spiking neuron circuit. The embodiment applies a first logical AND function to a first spike bit in the set of spike bits and a first synaptic weight of a first synapse in the set of synapses. The embodiment increments a membrane potential value associated with the spiking neuron circuit based on the applying. The embodiment determines that the membrane potential value associated with the spiking neuron circuit reached a learning threshold value. The embodiment then performs a Spike Time Dependent Plasticity (STDP) learning function based on the determination that the membrane potential value of the spiking neuron circuit reached the learning threshold value.