Improving Uncertainty Quantification in Deep Bayesian Active Learning

In many practical scenarios, labeling data can be highly costly, e.g., medical imaging, where expert radiologists are required, or legal document classification, which demands input from domain specialists. Active learning tackles this issue by iteratively selecting the most informative data points for labeling, with the goal of minimizing the number of labeled examples needed while maximizing model performance. Bayesian active learning is a specific approach to active learning that uses the Bayesian framework, particularly its treatment of uncertainty, to strategically select data points for labeling.

While Bayesian active learning has proven effective across various domains, such as natural language processing and image classification, it still faces several challenges. Many of these challenges stem from the uncertainty estimates produced by the active learning model. One issue is calibration, which arises by training on data that is not sampled from the true data distribution. Another issue arises with the use of Bayesian deep learning models: while they offer a strong inductive bias for a wide range of tasks and datasets, they often fail to provide reliable uncertainty estimates in the low-data regime, which is crucial for active learning.

In my research, I aim to investigate methods for improving the uncertainty quantification of active learning models, thereby enabling more efficient and reliable data selection. One potential approach is, in particular for the case of calibration, to make use of the LURE estimator discussed in Farquhar (2021) to correct the bias that results from training on a distribution that is different from our data distribution. Another potential approach is to consider the framework of active statistical inference proposed in Zrnic and Candes (2024). Regarding the Bayesian deep learning issue raised above, one approach is to make use of the recently developed linearized Laplace (Immer, 2021) approximation which benefits from the more reliable uncertainty estimates of a Gaussian process while having the inductive bias of a neural network.

Moreover, I also aim to investigate the application of active learning for de novo drug design. More specifically, I aim to investigate the use of active learning for identifying suitable ligands that can bind to a protein target of interest, where the candidate ligands can be generated by recent developments in using diffusion models to generate molecules, and where the active learning can be performed in latent space for increased computational and data efficiency (Bickford-Smith, 2024).