Research Projects

Trustworthy AI is a recent branch of research concerned with making AI tools (based on deep networks) useful in practice. The technical aspects of Trustworthy can be distilled into two major questions - (i) Can I use the given AI tool (usually a pre-trained network) for inference on a given samples - Called Out-Of-Distribution Detection, and (ii) Can the inference be trusted? - Called Calibration. In this project we aim to answer both these challenges using a quantile based approach. Several of the current approaches treat the questions (i) and (ii) as different questions and do not offer a common solution to both these problems. In this proposal we show that quantile representations can be used to solve both these problems at the same time. We illustrate it's working using toy examples and also provide theoretical grounding for the results. The main aim of the project is to scale the idea of computing quantile representations to very large networks and across wide range of tasks - Object Detection, Depth Estimation, Generative Models etc. The main challenges we aim to solve during this project are - (i) Theoretical extensions are required to extend the results from classification to more complex tasks such as object detection. (ii) Theoretical extensions to high dimensional quantiles, which are much more suitable for the multi-class classification as opposed to one-vs-rest approach. (iii) Designing efficient data-structures and algorithms for fast computation of quantile representations. The algorithms developed would allow for distributed training of large scale systems. The main outcome of this project would be a theoretically grounded framework to characterize the applicability of large networks for a given application. We hope to release a open-source library of the framework developed to allow extensions of our work as well as wide adoption of the proposed framework.