Embedded Files
About Me
About Me
I'm a PhD candidate at the Paul G. Allen School of Computer Science at the University of Washington , advised by Professor Byron Boots.
My research has involved a variety of machine learning methods and applications, the common thread across which has been to identify and utilize geometric structure when it exists. Currently, I'm designing metric-informed reinforcement learning algorithms that exploit geometric structure in MDPs to improve generalization and transfer.
Previously, I worked on understanding and applying quantum-inspired probabilistic models for sequential data. Recently, I incorporated geometric information about data spaces into sampling algorithms.
Current Projects
Current Projects
Research Intern at ChipStack:
LLMs for Chip Design
Research Intern at ChipStack:
LLMs for Chip Design
I am currently a Research Intern at Chipstack, where I am fine-tuning and optimizing large language models for chip design.
Behaviorally Aligned Representation Transfer
Behaviorally Aligned Representation Transfer
We are working on expanding our work on Behavioral Eigenmaps for representation transfer between reinforcement learning policies.
Publications
Publications
tl;dr Reinforcement learning from images is difficult because of high-dimensional observations. Behavioral distance based algorithms learn state-representations from images where behaviorally similar states are grouped together. This can help downstream policy learning. BeigeMap representations preserve local metric structure of such distances, instead of trying to match all distances globally. This drop-in modification improved policy performance in our experiments.
Adhikary, S., Li. A, & Boots, B. (2024). BeigeMaps: Behavioral Eigenmaps for Reinforcement Learning from Images. International Conference on Machine Learning (ICML).
How do we encode the geometric structure of behavioral metrics into representations for RL?
RLDM Travel Award! (Declined due to COVID-19)
tl;dr In multi-objective reinforcement learning, we often combine multiple policies to obtain a composite policy that is ultimately executed. When deciding how these policies should be combined, existing approaches generally opt for simple sums or products. We propose a framework to view the problem as finding centroids in distances spaces where policies are embedded.
Adhikary, S. and Boots, B. (2022). Modular Policy Composition with Policy Centroids. Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM)
How can we combine multiple modular policies to form a composite one?
Winner of best workshop paper award at R:SS 2021!
tl;dr In many applications, data often lies on a Riemannian manifold -- e.g. in robotics, orientation and inertia can be parameterized as points on a (hyper) sphere and the positive semi-definite cone. In such domains, standard sampling algorithms generally don't work out of the box, or work better with some adaption. We propose adaptations for the kernel herding sampling algorithm using geodesic kernels and Riemannian gradient descent.
Adhikary, S. and Boots, B. (2022). Sampling over Riemannian Manifolds using Kernel Herding. IEEE International Conference on Robotics and Automation (ICRA)
Adhikary, S., Thompson, J., and Boots, B., (2021). Sampling over Riemannian Manifolds with Kernel Herding. Robotics: Science and Systems (R:SS 2021) Workshop on Geometry and Topology in Robotics
How do we sample data using the kernel herding algorithm when the samples must lie on Riemannian manifolds?
tl;dr Many probabilistic graphical models have tried to expand the expressiveness of simple models like hidden Markov models (HMMs). Such models have been simultaneously proposed in different fields including stochastic processes, weighted automata, and quantum tensor networks. But it is not always clear how these models are related to one another. We prove equivalencies and build hierarchies of expressiveness between such models -- helping to bridge seemingly disparate models from these different fields.
Adhikary, S.*, Srinivasan, S.*, Miller, J., Rabusseau, G. and Boots, B., (2021). Quantum Tensor Networks, Stochastic Processes, and Weighted Automata. In International Conference on Artificial Intelligence and Statistics (AISTATS)
* indicates equal contribution
How are the probabilistic graphical models from stochastic processes, quantum tensor networks and weighted automata related?
tl;dr Joint probability distribution over sequences can be encoded as quantum tensor networks, which conform to certain constraints -- including trace preservation. When learning from data, we often want to define a constraint on model parameters locally (i.e. for a single block of the network). In this paper, we analyze the problem of defining such a local trace preservation constraint, and outline next steps in developing optimization algorithms.
Srinivasan, S., Adhikary S., Miller, J., Pokharel, B., Rabusseau, G. and Boots, B., (2021). Towards a Trace-Preserving Tensor Network Representation of Quantum Channels. Second Workshop on Quantum Tensor Networks in Machine Learning at NeurIPS.
Can we define local constraints on quantum tensor networks to ensure global trace preservation?
tl;dr Hidden Quantum Markov models (HQMMs) are generalizations of hidden Markov Models (HMMs) that utilize probabilistic states and dynamics from quantum mechanics. This generalization comes at the cost of new non-trivial constraints on model parameters. Using the fact that HQMM parameters lie on the Stiefel manifold, we propose an approach to learning them using Riemannian gradient descent. This not only yielded more accurate models than the prior approach, but was also much faster to train.
Adhikary, S.*, Srinivasan, S.*, Gordon, G., and Boots, B. (2020). Expressiveness and Learning of Hidden Quantum Markov Models. In International Conference on Artificial Intelligence and Statistics (AISTATS)
* indicates equal contribution
How do we learn hidden Quantum Markov models whose parameters lie on the Stiefel manifold?
Page updated
Google Sites
Report abuse