Machine Vision Training With Iterative Coreset Sub-Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine vision systems face memory constraints and lack of loss value output when using K-center greedy algorithms for anomaly detection, limiting their deployment on lower-end hardware.
Innovation Solution
A method involving iterative sub-sampling of training images into subsets, using a machine learning framework to extract patch-level features, and generating a coreset of features with a sub-sampling algorithm, while providing convergence values for real-time assessment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If K-center greedy algorithm is used for feature sub-sampling, then anomaly detection precision is improved, but memory consumption increases
Solution Approach 1:
The training dataset is divided into multiple batches, and the K-center greedy algorithm is applied iteratively to each batch separately. This segmentation allows the algorithm to process features in smaller chunks, reducing peak memory consumption while maintaining the overall precision of anomaly detection through cumulative coreset construction.
Solution Approach 2:
The algorithm pre-computes and stores the coreset features in memory before the actual anomaly detection process. By performing this sub-sampling action in advance during training, the system optimizes memory usage during inference while ensuring high detection precision through careful feature selection.
2Measurement precision
If K-center greedy algorithm is used for feature sub-sampling, then anomaly detection precision is improved, but loss value tracking becomes unavailable
Solution Approach 1:
The system introduces a feedback mechanism that computes and tracks loss values during the iterative K-center greedy sub-sampling process. By calculating the loss between selected coreset features and the full feature set at each iteration, the system provides visibility into training convergence while maintaining anomaly detection precision, allowing users to monitor the training process effectively.
3Quantity of substance
If iterative sub-sampling into subsets is implemented, then memory usage is reduced, but training time increases
Solution Approach 1:
The training process is segmented into multiple iterations over different batches of data. While this segmentation reduces memory usage by processing smaller subsets, the system optimizes training time by implementing efficient feature extraction and reuse across iterations, avoiding redundant computations and maintaining overall training efficiency.
Solution Approach 2:
The iterative sub-sampling process maintains continuity by accumulating coreset features across iterations rather than starting fresh each time. This continuous accumulation of useful feature information ensures that the training process remains efficient despite multiple passes, reducing the overall time penalty while maintaining low memory usage throughout the process.
Data Source
AI summary
Systems and methods are provided for training an imaging system for anomaly detection through a machine learning architecture combined with a local memory optimizing, iterative sub-sampling process. Training includes separating training images into subsets and iteratively feeding each subset to the machine learning architecture which extracts patch-level features in a feature space. A sub-sampling process generates a coreset from these extracted features, each iteration. Each iteration new extracted features and the existing coreset and are fed to the sub-sampling process which updates the coreset, iteratively until all training images are consumed. To aid optimization, each iteration a convergence value indicating coreset generation progress is determined for displaying status of the anomaly detection training.