Training-Free Neural Active Learning with Initialization-Robustness Guarantees
This work addresses the need for initialization robustness in safety-critical applications, offering an incremental improvement over existing neural active learning methods.
The paper tackles the problem of neural active learning by introducing a training-free criterion (EV-GP) that selects data points to ensure both good predictive performance and robustness against random parameter initializations, with empirical results showing it consistently outperforms baseline methods, especially in limited data or large batch scenarios.
Existing neural active learning algorithms have aimed to optimize the predictive performance of neural networks (NNs) by selecting data for labelling. However, other than a good predictive performance, being robust against random parameter initializations is also a crucial requirement in safety-critical applications. To this end, we introduce our expected variance with Gaussian processes (EV-GP) criterion for neural active learning, which is theoretically guaranteed to select data points which lead to trained NNs with both (a) good predictive performances and (b) initialization robustness. Importantly, our EV-GP criterion is training-free, i.e., it does not require any training of the NN during data selection, which makes it computationally efficient. We empirically demonstrate that our EV-GP criterion is highly correlated with both initialization robustness and generalization performance, and show that it consistently outperforms baseline methods in terms of both desiderata, especially in situations with limited initial data or large batch sizes.