Convolutional Neural Network Crossbar Circuit Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large-scale convolutional neural networks face challenges in performing convolution operations due to the physical limitations of crossbar circuits, such as IR drops and maximum current limitations, which restrict the number of input bars and output bars, hindering the use of dedicated hardware for efficient operations.
Innovation Solution
The implementation of a convolutional neural network that utilizes multiple crossbar circuits for convolution operations, with a merging layer to combine results from these circuits, allowing for efficient weight assignment and operation even when the number of input bars exceeds physical limits, thereby enabling the continuation of processes without degrading recognition performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single crossbar circuit is used for convolution operations, then the device structure is simple, but the number of input bars is limited by physical constraints (IR drops, maximum current)
Solution Approach 1:
The patent divides a single large crossbar circuit into multiple smaller crossbar circuits. Each crossbar circuit processes a subset of feature maps, and the results are merged through a merging layer. This segmentation allows the system to overcome the physical input bar limitations of individual crossbar circuits while maintaining the benefits of dedicated hardware acceleration for convolution operations.
2Quantity of substance
If multiple crossbar circuits are used to increase the number of input bars, then the physical limitations are overcome, but the device complexity increases
Solution Approach 1:
The patent introduces a merging layer that combines the output results from multiple crossbar circuits. This merging layer integrates the partial results from each crossbar circuit to produce the final convolution output, effectively combining the computational power of multiple circuits while managing the overall system complexity through a structured approach.
3Measurement precision
If the number of input bars is increased to handle large-scale convolutional neural networks, then the recognition performance can be maintained, but the current limitations (IR drops, maximum current) prevent further expansion
Solution Approach 1:
By segmenting the convolutional neural network into multiple crossbar circuits, each handling a portion of the feature maps, the system achieves the equivalent of having a larger number of input bars without being constrained by the physical limitations of a single crossbar circuit. This segmentation enables large-scale CNN processing while avoiding IR drops and current limitations.
4Productivity
If dedicated hardware crossbar circuits are used for efficient convolution operations, then the processing speed improves, but the physical constraints limit the scale of operations
Solution Approach 1:
The patent segments the convolutional neural network into multiple crossbar circuits that operate in parallel, each processing specific feature maps. This segmentation maintains the high processing speed benefits of dedicated hardware while enabling the system to handle larger-scale operations by distributing the computational load across multiple circuits.
Solution Approach 2:
The merging layer combines the results from multiple parallel crossbar circuits, integrating their computational outputs to produce the final convolution result. This merging mechanism enables the system to achieve the productivity of multiple dedicated hardware circuits working together while managing the overall hardware configuration through a unified structure.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This configuration effectively manages the physical constraints of crossbar circuits by distributing input bars across multiple circuits and merging their outputs, improving recognition performance and reducing error rates in classification tasks.
Implementation Method 1
The crossbar circuit performs a convolution operation in an analog region with respect to input data including the input signal by adding the input signals at each output bar
Data Source
AI summary
A convolutional neural network includes: convolution layers and a merging layer. At least one convolution layer includes a crossbar circuit having input bars, output bars and weight assignment elements that assign weights to input signals. The crossbar circuit performs a convolution operation in an analog region with respect to input data including the input signal by adding the input signals at each output bar. The input data includes feature maps. The crossbar circuit includes a first crossbar circuit for performing the convolution operation with respect to a part of the feature maps and a second crossbar circuit for performing the convolution operation with respect to another part of feature maps. The merging layer merges convolution operation results of the first and second crossbar circuits.