Efficient Leakage-Free Neural Architecture Search under Leave-One-Subject-Out Evaluation
Leave-One-Subject-Out (LOSO) evaluation makes Neural Architecture Search (NAS) computationally prohibitive, scaling as O(N^2) due to N independent architecture searches A leakage-free, block-based NAS approach is proposed that shares search runs across subjects, dramatically reducing computational cost On the BioVid Heat Pain dataset, the method improved mean accuracy from 82.79% to 83.39% while cutting model parameters by up to 99.2% The approach eliminates data leakage while maintaining or imp
Analysis
TL;DR
- Leave-One-Subject-Out (LOSO) evaluation makes Neural Architecture Search (NAS) computationally prohibitive, scaling as O(N^2) due to N independent architecture searches
- A leakage-free, block-based NAS approach is proposed that shares search runs across subjects, dramatically reducing computational cost
- On the BioVid Heat Pain dataset, the method improved mean accuracy from 82.79% to 83.39% while cutting model parameters by up to 99.2%
- The approach eliminates data leakage while maintaining or improving generalization performance in subject-based classification tasks
- This work addresses a critical scalability bottleneck in personalized/subject-specific deep learning model development
Why It Matters
This research directly addresses one of the most pressing computational challenges in subject-specific machine learning: the prohibitive cost of performing NAS under rigorous cross-subject evaluation protocols. For practitioners working in healthcare, affective computing, and biometric domains where subject-independent generalization is paramount, this method offers a practical path to efficient architecture discovery without sacrificing evaluation integrity.
Technical Details
- Problem framing: LOSO-NAS requires N independent architecture searches (one per held-out subject), leading to O(N^2) training cost under approximately linear training scaling assumptions
- Proposed solution: A block-based NAS approach that shares search runs across subjects while maintaining strict leakage-free guarantees, avoiding information bleed between training and validation folds
- Benchmark: Evaluated on the BioVid Heat Pain dataset, a well-known physiological signal dataset for pain classification across subjects
- Results: Mean accuracy improved from 82.79% to 83.39%, with parameter reduction of up to 99.2%, indicating discovery of significantly more compact architectures
- Domain: Machine Learning (cs.LG) and Artificial Intelligence (cs.AI), arXiv:2609.09433
Industry Insight
- Subject-specific NAS is increasingly relevant as personalized AI applications grow in healthcare and human-computer interaction; this work provides a scalable blueprint for reducing search costs by an order of magnitude
- The 99.2% parameter reduction demonstrates that shared NAS can discover far more efficient architectures than subject-specific searches, suggesting practitioners should prioritize cross-subject sharing strategies in resource-constrained deployments
- The leakage-free guarantee is critical for clinical and regulated applications where evaluation integrity directly impacts safety certifications; this methodology should become a standard reference for subject-wise ML pipelines
Disclaimer: The above content is generated by AI and is for reference only.