Preprints for all papers are at arXiv
Active Learning for Neural Network models
The current paradigm of deep learning asks for large datasets, often labelled by humans, to train modern neural network models. Can we obtain a high quality model without paying such a steep labelling cost? Active Learning aims to achieve this by only labelling points that are determined to be most informative. In recent work we have extended the classical work of Chernoff on the sequential design of experiments for active hypothesis testing to active linear regression. In this research thrust, we aim to extend these ideas to the challenging domain of neural networks.
Subhojyoti Mukherjee*, Ardhendu Tripathy*, Robert Nowak (*Equal contribution)
Sequential Learning for Real-world Physical Processes
Sequential learning has been used by agents to improve their policy over the course of interaction with the environment. Common examples of these agents include recommendation systems and game-playing agents. In this research thrust we aim to employ methods developed in sequential learning literature to real-world processes. We choose the case study of deficit irrigation in agricultural fields, and will develop a framework to automatically optimize irrigation amounts.
Aligning Algorithmic Decision-makers with Humans
With the help of a good enough simulation model, or with enough interactions with the environment, an agent can improve its performance as measured by a utility function. However, the agent utility is most likely be mismatched to the utility a human operator may have in mind. In this research thrust, using novel means of visualization and surrogate modeling, we aim to align an algorithmic decision-maker with the interpretable attributes of the human operator.
Blake Mason, Lalit Jain, Ardhendu Tripathy, and Robert Nowak
Sumeet Katariya*, Ardhendu Tripathy*, Robert Nowak (*Equal contribution)
Blake Mason*, Ardhendu Tripathy*, Robert Nowak (*Equal contribution)