Learning Device Selects Image Data by Confidence to Reduce Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning for image classification requires large memory sizes due to the large data size of images, and existing methods either focus on low-confidence data or exclude data from learning, leading to inefficient memory usage and learning processes.
Innovation Solution
A learning device and method that selects image data based on classification confidences, stores the selected data, performs learning, and updates confidences when a predetermined condition is met, reducing memory requirements and improving learning efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all image data is stored in memory for learning, then learning accuracy is improved, but memory size requirement increases significantly
Solution Approach 1:
The patent segments the image data into batches that can be processed sequentially rather than loading all data at once. The learning unit processes multiple batches of image data, where each batch is a subset of the total training data, allowing memory-efficient processing while maintaining learning accuracy through multiple passes over the data.
Solution Approach 2:
The patent implements continuous learning by repeatedly processing multiple batches of image data through the learning unit. The learning model is continuously updated across batches and iterations, ensuring that useful learning action continues without interruption while managing memory constraints through batched processing.
2Quantity of substance
If image data is processed in batches, then memory usage is reduced, but learning time increases
Solution Approach 1:
The patent applies periodic action by processing image data in regular batches with predetermined batch sizes. The learning unit periodically processes each batch, updates the learning model, and moves to the next batch, creating a rhythmic processing pattern that balances memory usage with learning efficiency through structured iteration.
3Productivity
If learning data is selected based on classification confidence, then learning efficiency is improved, but data selection complexity increases
Solution Approach 1:
The patent changes the parameter of data selection from random or uniform sampling to confidence-based sampling. The image selection unit uses classification confidence values as a parameter to identify and select uncertain image data for learning, transforming the selection criterion to focus on data that would benefit most from learning while maintaining manageable complexity through confidence thresholding.
Data Source
AI summary
Provided is a learning device that includes at least one processing device configured to select image data as learning data based on classification confidences from a plurality of image data. Each of the classification confidences indicates a likelihood of accuracy of classification for a respective one of the plurality of image data. The at least one processing device is configured to: store the selected image data into a storage; perform learning using the stored image data and generating a learning model; perform classification of the image data used for the learning by using the learning model and update a classification confidence for the classified image data; and update the classification confidences each associated with a respective one of the plurality of image data when a count of image data on which the classification has been performed satisfies a predetermined condition.