Neuromorphic Device Multi-Channel Binarization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neuromorphic devices face challenges in efficiently processing and classifying input data due to limitations in binary feature map generation and weight storage, which affects the accuracy and power efficiency of neural network operations.
Innovation Solution
The method involves generating multiple binary feature maps based on various thresholds, binarizing pixel values, and using a crossbar array circuitry for multiplications with stored weight values, followed by selective merging of output values to produce an output feature map, which can be used for further processing in neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If binary feature maps are generated using single-channel binarization, then device complexity is reduced, but classification accuracy deteriorates
Solution Approach 1:
The patent divides the single binarization channel into multiple parallel channels, each processing feature maps with different threshold values. This segmentation allows the system to maintain low device complexity while improving classification accuracy by capturing diverse feature representations simultaneously.
Solution Approach 2:
The patent extends the binarization process from a single-dimensional approach to a multi-dimensional approach by introducing multiple threshold values across different channels. This dimensional expansion enables richer feature extraction without proportionally increasing device complexity.
2Measurement precision
If multiple binary feature maps are generated with multiple thresholds, then classification accuracy is improved, but power consumption increases
Solution Approach 1:
The patent extracts only the essential computational operations needed for multi-threshold binarization, implementing them efficiently in the crossbar array circuitry. By taking out and optimizing only the critical path operations, the system achieves improved accuracy without proportionally increasing power consumption.
Solution Approach 2:
The patent uses multiple binary feature maps as parallel copies of the input data processed through different threshold channels. This copying approach enables accurate feature extraction while allowing power-efficient operations since binary data requires less energy to process than high-precision data.
3Productivity
If weight values are stored in synaptic circuits of crossbar array circuitry, then multiplication efficiency is improved, but device complexity increases
Solution Approach 1:
The patent merges the weight storage function and multiplication function into a single integrated structure—the crossbar array circuitry with synaptic circuits. This merging eliminates the need for separate memory and processing units, improving multiplication efficiency while keeping device complexity manageable through functional integration.
Solution Approach 2:
The synaptic circuits in the crossbar array circuitry serve multiple functions: storing weight values, performing multiplication operations, and supporting different threshold-based binarization channels. This multi-functionality improves productivity without requiring separate dedicated circuits for each function.
Data Source
AI summary
A neuromorphic device and method are provided. A neuromorphic method includes generating a plurality of binary feature maps by multi-channel, based on a plurality of thresholds, binarizing pixel values of an input feature map, providing pixel values of each of the plurality of binary feature maps as input values to a crossbar array circuitry, storing weight values of a machine model in respective synaptic circuits included in the crossbar array circuitry, generating output values of the crossbar array circuitry for the plurality of binary feature maps by implementing multiplications respectively between each of a plurality of the input values and corresponding weight values stored in the synaptic circuits, and generating pixel values of an output feature map by selectively merging the output values.