Auxiliary objective improves generalization performance but reduces model specification for low-data neuroimaging-based brain age prediction
Track: Extended Abstract Track
Keywords: Underspecification, Structural magnetic resonance imaging, Generalization, Neural network, Brain Age
Abstract: Data scarcity and underspecification are 2 common issues in machine learning for healthcare. Data scarcity impedes the performance and generalizability of neural networks. Underspecification (where the training process can produce many different models that achieve the same train/test performance but represent different functions) may lead to issues with generalization and poor model behavior in deployment settings. In this work, we add an auxiliary objective to a brain age prediction model that significantly improves model performance and generalization in low-data regimes. We evaluate the impact of the auxiliary objective on model specification and particularly quantify how random variations in the training process affect a model's representations and predictions. Our results show that while the auxiliary objective enhances generalization and performance, especially in data-limited settings, it also reduces model specification. These findings underscore the trade-off between improving generalization with added constraints such as auxiliary losses, and their reduction in model specification in low-data neuroimaging applications.
Submission Number: 62
Loading