Our research group is advancing a data-centric view of Artificial Intelligence (AI): we develop principled approaches that make the usage of data in modern AI models more efficient. The developed algorithms improve efficiency by using sample informativeness to guide the data collection process or by learning a data representation that improves the performance of AI models. We achieve this under two research themes that include theoretical analysis and empirical validation of our developed methods.
- Sequential decision-making
- Representation learning
Sequential decision-making
We are developing algorithms that efficiently solve novel pure exploration problems in multi-armed bandit environments. These problems are motivated by applications ranging from recommender systems and automated drug discovery to the tuning of parameters in engineering design simulators. These problems require us to utilize the latest experimental observations in the search for good candidates within a large search space. Our research has obtained near optimal algorithms for finding the subset of all epsilon-good arms [1] and for finding the subset defined by the largest gap between consecutive arm means [2].
A good algorithm incorporates the structure of the environment and the problem objective in its decision-making process, which must strike an optimal balance between exploration (improving uncertain estimates) and exploitation (using current estimates to minimize cost). Our work in [3] has further developed this principle by analyzing the cost of exploration and exploitation in the non-asymptotic, fixed confidence regime. Our ongoing work aims to elucidate the role of structure in multi-armed bandit environments and develop active sampling strategies for AI model training. Our research will enable practical, data-efficient search and training procedures for predictive AI models and broaden understanding of sequential decision-making algorithms.
Representation learning
We are developing methods that enable AI models to learn representations of data that are beneficial for a desired objective. For example, our research has used generative adversarial networks to learn representations of data that are invariant to a sensitive attribute [4]. The resulting “sanitized” version of the dataset can be used in privacy-preserving data analyses.
Beyond privacy, the high-dimensional nature of modern datasets makes computing relevant features or statistics a challenging representation learning problem. The computational costs can be reduced by relying on approximate or noisy computation. However, despite noisy computations, we can learn accurate representations of the data using bandit optimization techniques. Bandit optimization techniques can also be used when the noise is due to the variability in human judgments. An example is crowdsourcing; our work in [5] has learned nearest neighbors in a dataset using distance values inferred from human preferences. Our ongoing work is developing representations that reduce the number of function evaluations needed in derivative-free optimization of black-box functions.
[1] Mason, B., Jain, L., Tripathy, A., & Nowak, R. (2020). Finding all e-good arms in stochastic bandits. Advances in Neural Information Processing Systems, 33, 20707-20718.
[2] Katariya, S.*, Tripathy, A.*, & Nowak, R. (2019). Maxgap bandit: Adaptive algorithms for approximate ranking. Advances in Neural Information Processing Systems, 32. (*Equal contribution)
[3] Mukherjee, S.*, Tripathy, A.*, & Nowak, R. (2022). Chernoff sampling for active testing and extension to active regression. In International Conference on Artificial Intelligence and Statistics (pp. 7384-7432). PMLR. (*Equal contribution)
[4] Tripathy, A., Wang, Y., & Ishwar, P. (2019). Privacy-preserving adversarial networks. In 2019 57th Annual Allerton Conference on Communication, Control, and Computing (Allerton) (pp. 495-505). IEEE.
[5] Mason, B.*, Tripathy, A.*, & Nowak, R. (2019). Learning nearest neighbor graphs from noisy distance samples. Advances in Neural Information Processing Systems, 32. (*Equal contribution)