Neural Network Composition Conversion for Embedded Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network implementations on embedded devices face a trade-off between processing performance and recognition accuracy due to the need for reduced computation, where decreasing multiply-accumulate operations to save resources often results in lower accuracy.
Innovation Solution
A composition conversion apparatus that analyzes and converts layer parameters of neural networks to increase computation without deteriorating processing performance, specifically by adjusting the number of input/output edges in each layer to enhance the neural network's computation amount while maintaining performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the computation amount is reduced to enable real-time processing on inexpensive embedded devices, then processing performance is improved, but recognition accuracy deteriorates
Solution Approach 1:
The patent changes the parameters of the neural network layers by converting layer parameters to increase the computation amount. Specifically, it adjusts the number of input/output edges in each layer and modifies weight matrices to have more elements, thereby increasing the overall computation amount while maintaining processing performance on embedded devices
Solution Approach 2:
Instead of reducing computation to improve performance, the patent inverts the approach by increasing computation amount while still achieving performance improvement. This is done by converting layer parameters to create a neural network structure that performs more computations but is optimized for efficient execution on embedded devices
2Measurement precision
If the computation amount is increased to improve recognition accuracy, then recognition accuracy is improved, but processing performance deteriorates
Solution Approach 1:
The patent modifies layer parameters including the number of input/output edges and weight matrix dimensions to increase computation amount for better accuracy, while the conversion process optimizes these parameters to maintain processing performance on resource-constrained embedded devices
3Ease of manufacture
If the number of multiply-accumulate operations is reduced to save resources on embedded devices, then device cost is reduced, but recognition accuracy deteriorates
Solution Approach 1:
The patent converts layer parameters to increase the computation amount by adjusting the number of input/output edges and weight matrix elements. This creates a neural network that requires more computations for better accuracy while still being executable on inexpensive embedded devices through efficient parameter configuration
Data Source
AI summary
An analysis unit (110) analyzes a composition of a neural network composed of a plurality of layers and acquires a layer parameter (210) indicating an attribute of each of the plurality of layers. A conversion unit (120) converts the layer parameter (210) in such a way that processing performance of a circuit which executes an operation of the neural network does not deteriorate and a computation amount of the neural network increases. The conversion unit (120) increases the computation amount by treating the numbers of input/output edges in each of the plurality of layers as the layer parameter and increasing the numbers of input/output edges.