Neural Network Processor Using Multi-Valued Basis Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep Convolutional Neural Networks (DCNN) require significant processing time and memory due to high calculation and parameter demands, making it difficult to incorporate trained models into low-spec devices like embedded devices and mobile devices without re-training.
Innovation Solution
A neural network processor that performs multi-valued neural network processing using a control circuitry, quantization processing, and dot product processing, reducing the number of dimensions in scaling coefficient vectors through the use of multi-valued basis matrices, allowing efficient high-accuracy calculation without re-training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If DCNN with many intermediate layers is used to improve recognition accuracy, then recognition accuracy is improved, but processing time and model size are increased
Solution Approach 1:
The patent applies parameter changes by transforming the basis matrix from binary values to multi-valued data, and by adjusting the quantization width dynamically. This allows the system to maintain high recognition accuracy while reducing the number of dimensions in scaling coefficient vectors, thereby decreasing processing time and model size without sacrificing accuracy.
Solution Approach 2:
The patent introduces dimensionality change by reducing the number of dimensions in scaling coefficient vectors through the use of multi-valued basis matrices. This dimensional reduction allows the neural network to process information more efficiently, achieving faster processing times while maintaining the necessary recognition accuracy through the enriched multi-valued representation.
2Measurement precision
If DCNN with many intermediate layers is used to improve recognition accuracy, then recognition accuracy is improved, but model size is increased
Solution Approach 1:
The patent changes key parameters including the basis matrix values (from binary to multi-valued) and quantization width. These parameter changes enable a more compact model representation that reduces model size while preserving recognition accuracy through the increased information density of multi-valued elements.
Solution Approach 2:
By reducing the number of dimensions in scaling coefficient vectors through multi-valued basis matrices, the patent achieves a more compact model structure. This dimensional reduction directly decreases model size while the multi-valued representation maintains the necessary complexity for high accuracy recognition.
3Measurement precision
If a trained model from large-scale system is incorporated into low-spec devices, then recognition performance is achieved, but re-training is required which increases complexity
Solution Approach 1:
The patent enables direct copying of trained models from large-scale systems to low-spec devices by using a representation format (multi-valued basis matrices with reduced-dimensional scaling coefficient vectors) that is both space-efficient and computationally lightweight. This eliminates the need for re-training while preserving recognition performance.
Solution Approach 2:
By transforming the model parameters into a multi-valued representation with reduced dimensions, the patent creates a version of the trained model that is optimized for low-spec devices. This parameter transformation allows direct deployment without re-training, reducing complexity while maintaining performance.
4Productivity
If binary basis matrix is used to reduce calculation amount, then processing speed is improved, but approximation accuracy is insufficient
Solution Approach 1:
The patent changes the basis matrix from binary to multi-valued, which increases the information content per element. This parameter change allows for more accurate approximation of the original weights while still enabling efficient processing, as the multi-valued representation captures nuances that binary cannot, without requiring proportionally more computation.
Solution Approach 2:
The patent combines dimensionality reduction in the scaling coefficient vectors with multi-valued basis matrices. This approach maintains processing speed by keeping the overall structure compact while improving approximation accuracy through the richer multi-valued representation, achieving a better balance than binary matrices.
Data Source
AI summary
Provided is a processor for a neural network whose high-performance compact model can be incorporated into low-spec devices such as embedded devices or mobile devices without requiring re-training. The processor for a neural network, which uses a multi-valued basis matrix, widens the range of integer values that can be taken by each element of the multi-valued basis matrix; thus, the number of dimensions (the number of elements) of a scaling coefficient vector is reduced accordingly. The elements of the scaling coefficient vector are real numbers, and thus reducing the amount of processing of real number calculation processing allows for reducing the number of dimensions (the number of elements) of the scaling coefficient vector. As a result, this neural network processor significantly reduces the amount of calculation processing while ensuring the calculation accuracy when performing matrix calculation processing using the binary basis matrix.