Neural Network Data Processing Device Using Segmented Binary Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing devices for neural networks face challenges in achieving high recognition accuracy while minimizing memory capacity, especially for embedded devices with limited resources.
Innovation Solution
A data processing device is designed with an input data determining unit, a storage unit for coefficients and coefficient address information, and an arithmetic unit that executes cumulative additions based on binarized input data and stored coefficients, allowing for efficient use of memory and arithmetic resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point weight parameters are used in neural network inference processing, then recognition accuracy is improved, but memory capacity requirement increases
Solution Approach 1:
The weight parameter is segmented into multiple components: basis information (binary data), coefficient information, and offset information. This segmentation allows the heavy floating-point weight parameter to be represented using lighter binary basis data combined with coefficient and offset values, reducing memory capacity requirements while maintaining recognition accuracy through the reconstructed weight parameter W = Σ(Ci × Bi) + Offset
Solution Approach 2:
The invention changes the parameter representation from direct floating-point weights to a transformed representation using binary basis information, coefficients, and offsets. This parameter transformation enables the same neural network functionality to be achieved with reduced memory capacity by storing binary basis data and coefficient information instead of full floating-point weight parameters
2Quantity of substance
If binary data is used to approximate weight parameters, then memory capacity is reduced, but the distribution of expressible values becomes limited and symmetrical
Solution Approach 1:
The invention introduces asymmetric offset information to compensate for the inherent symmetry of binary basis data combinations. By adding an offset component (Offset = μ - Σ(Ci × Bi)), the system can represent weight parameters with asymmetric distributions that match the actual data distribution, thereby improving adaptability while maintaining reduced memory capacity
Solution Approach 2:
The invention uses coefficient information as an intermediary between binary basis data and the final weight parameter representation. The coefficients act as mediators that scale the binary basis values, enabling a broader range of expressible values while still using compact binary representations, thus bridging the gap between memory efficiency and value distribution versatility
3Power
If floating-point addition is used instead of multiple accumulation, then computation resources are reduced, but recognition accuracy may decrease
Solution Approach 1:
The invention performs preliminary transformation of weight parameters into binary basis information and coefficient information before the inference processing. This preliminary action allows the subsequent computation to use simpler operations (addition of coefficients) instead of complex floating-point multiplications, reducing computation resources while maintaining recognition accuracy through the pre-computed basis representation
Data Source
AI summary
A data processing device includes: an input data determining unit configured to determine whether or not each of binarized input data is a predetermined value; a storage unit configured to store a plurality of coefficients and coefficient address information including information related to coefficient addresses where the plurality of coefficients are stored; a control unit configured to read the coefficient address from the storage unit based on a determination result of the input data determining unit and read the coefficient from the storage unit based on the coefficient address; and an arithmetic unit configured to execute an arithmetic operation related to the coefficient acquired by the control unit.