Tensor Image Mapping Device Using Threshold-Based Weight Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing neural network architectures, particularly in convolution and fully connected layers, face increased computational costs due to multiplication operations, which can be inefficient and costly in terms of hardware requirements.
Innovation Solution
A computing device that approximates calculations in convolution and fully connected layers by selecting and adding weight values based on threshold comparisons, reducing the need for multiplication and thereby decreasing computational complexity and hardware costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiplication operations are used in convolution and fully connected layers, then calculation precision is maintained, but computational cost and hardware requirements increase
Solution Approach 1:
The patent changes the computational operation from multiplication to addition by modifying the calculation formula. Instead of computing y(Ox, Oy) = f(Σw(i, j) × x(Ox+i, Oy+j)), it uses y(Ox, Oy) = f(Σselect(i, j, x(Ox+i, Oy+j)) + b), where weight values are selected based on threshold comparisons of input pixel values. This parameter change eliminates multiplication operations while maintaining approximation accuracy.
Solution Approach 2:
The patent replaces expensive multiplication operations with cheaper addition operations and threshold comparisons. The weight values are selectively applied based on simple threshold checks rather than being multiplied, significantly reducing the computational resources and hardware complexity required while maintaining acceptable calculation precision.
2Reliability
If multiplication operations are performed in neural network layers, then accurate pattern recognition is achieved, but processing time and computational load increase
Solution Approach 1:
The patent transforms the computational process by replacing multiplication with addition operations combined with threshold-based weight selection. This parameter change in the calculation methodology reduces computational load and processing time while maintaining sufficient accuracy for pattern recognition tasks through the selective application of weight values based on input characteristics.
Solution Approach 2:
The patent performs preliminary threshold comparisons to determine which weight values should be applied, replacing the need for actual multiplication operations. By pre-determining weight selection based on threshold checks of input pixel values, the system avoids computationally expensive multiplication while maintaining recognition accuracy.
3Adaptability or versatility
If standard convolution layer calculations are used, then neural network functionality is maintained, but computational cost increases
Solution Approach 1:
The patent modifies the computational parameters of the convolution layer by replacing multiplication operations with addition operations and threshold-based weight selection. This changes the energy consumption profile from high-cost multiplication to low-cost addition and comparison operations, significantly reducing computational cost while preserving the essential functionality of the neural network layer.
Data Source
AI summary
According to an embodiment, a computing device includes a processing circuitry. The processing circuitry receives an input of tensor data. The processing circuitry sets a window in the tensor data. The processing circuitry compares, for each pair of coordinates in the tensor data within the window, a pixel value at the pair of coordinates with one or more thresholds, and selects a weight value corresponding to a comparison result. The processing circuitry adds the weight values selected for the respective pairs of coordinates to obtain a cumulative value. The processing circuitry derives a value based at least in part on the cumulative value.