Robust Fine-Tuning of Deep Neural Networks with Hessian-based Generalization Guarantees
This work addresses the problem of overfitting during fine-tuning for practitioners in machine learning, offering a novel theoretical framework and algorithm, though it is incremental in building on existing generalization measures.
The paper tackles overfitting in fine-tuning pretrained deep neural networks, particularly on small or noisy datasets, by proposing a Hessian-based distance measure derived from PAC-Bayesian analysis that correlates with generalization gaps and provides theoretical guarantees, and it introduces an algorithm for label noise that shows gains over prior methods in image classification tasks.
We consider fine-tuning a pretrained deep neural network on a target task. We study the generalization properties of fine-tuning to understand the problem of overfitting, which has often been observed (e.g., when the target dataset is small or when the training labels are noisy). Existing generalization measures for deep networks depend on notions such as distance from the initialization (i.e., the pretrained network) of the fine-tuned model and noise stability properties of deep networks. This paper identifies a Hessian-based distance measure through PAC-Bayesian analysis, which is shown to correlate well with observed generalization gaps of fine-tuned models. Theoretically, we prove Hessian distance-based generalization bounds for fine-tuned models. We also describe an extended study of fine-tuning against label noise, where overfitting remains a critical problem. We present an algorithm and a generalization error guarantee for this algorithm under a class conditional independent noise model. Empirically, we observe that the Hessian-based distance measure can match the scale of the observed generalization gap of fine-tuned models in practice. We also test our algorithm on several image classification tasks with noisy training labels, showing gains over prior methods and decreases in the Hessian distance measure of the fine-tuned model.