Shashank Sharma

Hierarchical Reinforcement Learning For Natural Language

Project Summary

Hierarchical models function by estimating long-term and short-term objectives, close to how Human Language works using Incremental Procedural Grammar. Conveying a thought involves story-telling that takes a prior context through a set of transformations to a goal context. The transformations are generated on the fly and try to achieve the goal context, which can also be updated simultaneously. These transformations map to short sequences of words that get interwoven together and yield a coherent expression of thought. Additionally, since HRL models define long- and short-term objectives ahead of time, they offer full transparency and predictability.

 

Research Interests

Natural Language.

Reinforcement learning.

Cognitive science.

Background

Bachelors (B.Tech) in Mechanical Engineering from IIT Kanpur.

Worked with Google for the application of deep learning models to remote sensing imagery.

Supervisors

Dr. Marina De Vos

Dr. Vinay Namboodiri