Efficient neural network design improves subject-based predictions
Efficient Leakage-Free Neural Architecture Search under Leave-One-Subject-Out Evaluation
Machine LearningArtificial Intelligence
Summary
Predicting information for one person when you have data from many others can be slow and costly when designing neural networks. The authors developed a method to speed this up by sharing parts of the design process between people, avoiding mistakes that mix training and testing data. Their method improved accuracy slightly and greatly reduced the size of the networks tested on a pain detection dataset.
What this means in practice
- •For biomedical data scientists: Design smaller, more accurate neural networks for subject-specific biomedical signal classification using shared architecture search.
- •For wearable device engineers: Create efficient models for personalized health monitoring that reduce computational needs and maintain accuracy across users.
Tested on one dataset.
Authors
Heinke Hihn
Abstract
Leave-One-Subject-Out (LOSO) evaluation estimates generalisation performance for subject-based classification but makes Neural Architecture Search (NAS) computationally expensive because a fully nested implementation requires N independent architecture searches and, assuming approximately linear training cost, scales as O(N^2). We propose a leakage-free, block-based approach that shares NAS runs across subjects. On the BioVid Heat Pain dataset, our approach increased the mean accuracy from 82.79% to 83.39% while reducing the number of parameters by up to 99.2%.