Neural Network Compression Using Influence-Based Training Image Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural network compression methods for autonomous driving face challenges in maintaining recognition accuracy while reducing the arithmetic operation amount, leading to prolonged compression design periods due to the inclusion of harmful training images with noise or limited generalizability.
Innovation Solution
A neural network compression device that classifies training data into valid and invalid sets based on influence values, allowing for targeted compression and re-training using only the valid data to improve recognition accuracy and shorten the compression design period.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If conventional compression methods are used to reduce arithmetic operation amount, then the compression rate is improved, but recognition accuracy deteriorates due to inclusion of harmful training images
Solution Approach 1:
The training data set is segmented into valid training data and invalid training data based on influence values. This segmentation allows the compression process to selectively use only valid data, preventing harmful images from degrading recognition accuracy while still achieving arithmetic operation reduction through compression.
Solution Approach 2:
Invalid training data (harmful images with noise or limited generalizability) is extracted and removed from the training data set before compression. This extraction ensures that only beneficial training data is used in the compression process, maintaining recognition accuracy while reducing arithmetic operations.
2Manufacturing precision
If all training data including harmful images are used for compression design, then the training completeness is improved, but the compression design period is prolonged
Solution Approach 1:
The training data set is pre-processed to classify and remove invalid training data before the compression design process begins. This preliminary action reduces the volume of data that needs to be processed during compression design, shortening the design period while maintaining completeness by ensuring all valid data is retained.
Solution Approach 2:
Invalid training data is extracted and removed from the complete training data set. This extraction reduces the computational burden and time required for compression design while preserving all necessary valid training data, achieving both completeness and efficiency.
3Speed
If compression is performed without data classification, then the processing speed is improved, but the quality of compression results deteriorates
Solution Approach 1:
The training data is segmented into valid and invalid portions using influence value calculation. This segmentation enables the compression process to focus only on valid data, improving result quality while maintaining processing efficiency by avoiding unnecessary processing of harmful images.
Solution Approach 2:
Different quality standards are applied to different portions of the training data. Valid training data receives full processing attention to ensure high compression quality, while invalid data is identified and excluded. This local quality approach optimizes overall compression result quality without sacrificing processing speed.
Data Source
AI summary
When it is assumed that a large-scale Deep Neural Network for autonomous driving applied compression, there are problems of a decrease in recognition accuracy of a post-compression Neural Network (NN) model and an increase in a compression design period, due to a large number of harmful or unnecessary training images (invalid training images). A training image selection unit B100 calculates an influence value on an inference, and generates an indexed training image set 1004-1 necessary for an NN compression design, by using the influence value. A neural network compression unit P200 notified of the result via a memory P300 compresses the NN.