Turbo Online One-class Learning SVM for Fast Batch Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current support vector machines (SVMs) for one-class learning are computationally intensive due to the need to recalculate support vectors based on new and old primary data, making them inefficient for rapid online learning, especially in large-scale problems where storing all data is unfeasible.
Innovation Solution
The Turbo Online One-class Learning (TOOL) method employs a true divide-and-conquer decomposition approach that combines solutions of subproblems without recomputing previously defined support vectors, treating new data as a dependent subproblem to create low-dimensional quadratic programming problems, allowing for faster training and classification without relying on previous data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current SVM methods recalculate support vectors based on new and old primary data, then classification accuracy is maintained, but computational complexity increases and training speed decreases
Solution Approach 1:
The patent divides the training data into batches and processes them sequentially. Instead of recalculating support vectors from all data, it segments the problem into: (1) using previously computed support vectors from old data, and (2) computing corrections only from new data batches. This segmentation reduces computational complexity while maintaining classification accuracy.
Solution Approach 2:
The patent performs preliminary computation of support vectors from old data batches before new data arrives. These pre-computed support vectors are then reused and updated incrementally with new data, avoiding the need to recalculate from scratch and reducing overall computational complexity.
2Measurement precision
If current SVM methods store all primary data for recalibration, then learning accuracy is improved, but memory requirements and storage costs increase
Solution Approach 1:
The patent extracts and retains only the essential information from old data in the form of computed support vectors and their parameters. Instead of storing all primary training data, it extracts the critical components needed for classification and discards redundant data, significantly reducing memory storage requirements.
Solution Approach 2:
The patent uses copies of support vector information (parameters, coefficients) rather than storing original primary data. These compact representations contain the essential learning outcomes and can be reused multiple times without requiring access to the full original dataset.
3Reliability
If current SVM methods iterate over all training data multiple times, then convergence to optimal solution is achieved, but training time increases
Solution Approach 1:
The patent implements dynamic incremental learning where the model adapts to new data batches sequentially without requiring multiple passes over the entire dataset. The support vectors are updated dynamically as new data arrives, achieving convergence through incremental adaptation rather than repeated full-data iterations.
Solution Approach 2:
The patent maintains continuous learning by processing data batches sequentially without interruption. Each batch contributes incrementally to the final model through continuous updates of support vectors, eliminating the need to stop and restart multiple times over the complete dataset.
Data Source
AI summary
Methods for one-class learning using support vector machines from a plurality of data batches are provided. A first support vector machine is learned from the plurality of data batches by a processor. A new data batch is received by the processor and is classified by the first support vector machine. If a non-zero loss classification occurs a new support vector machine is trained using the first support vector machine and the new data batch only. Data batches can be discarded if they are represented by the current support vector machine or after being used for training an updated support vector machine. Weighing factors applied to update the first support vector machine depend upon a parameter which is optimized iteratively. Support vectors do not need to be recalculated. A classifier is learned in a number of stages equal to the number of data batches processed on-line.