Sequential Tests for Large-Scale Learning.

Area of Science:

• Machine Learning
• Statistical Inference
• Computational Statistics

Background:

• Traditional algorithms struggle with large datasets, requiring significant computational resources.
• Efficient processing of big data is crucial for modern machine learning and statistical inference.

Purpose of the Study:

• To develop adaptive subsampling algorithms for big data processing.
• To enhance the efficiency and accuracy of learning and inference algorithms.
• To provide statistical control over the subsampling process.

Main Methods:

• Introduction of algorithms utilizing sequential hypothesis tests for adaptive data subset selection.
• Application of these methods to learning by optimization, assessing update direction probabilities.
• Integration with posterior inference via Markov chain Monte Carlo, evaluating sample acceptance/rejection probabilities.

Main Results:

• Experimental evaluation on diverse models and datasets demonstrates the effectiveness of the proposed algorithms.
• The statistical properties of subsampling are shown to effectively control learning and inference performance.
• Adaptive subsampling offers a viable approach to managing computational complexity in big data scenarios.

Conclusions:

• Adaptive subsampling using sequential hypothesis tests is a promising strategy for big data challenges.
• The developed algorithms offer a tunable trade-off between computational efficiency and statistical accuracy.
• This approach has broad applicability in machine learning and statistical inference tasks.