Pratyusha Sharma
MIT
pratyuss@mit.edu
Bio
Pratyusha Sharma is a Ph.D. student at MIT, advised by Prof. Antonio Torralba and Prof. Jacob Andreas. She works on developing algorithms to understand the structure of solutions artificial neural networks implement, understanding the complexity and structure of naturally arising animal communication systems in the wild, and how language and natural-language-like structures can support effective reasoning and planning in embodied agents and robots. Her research has been published in interdisciplinary journals like Nature Communications and in top academic conferences across robotics, natural language processing, machine learning and marine biology. Her research has also been featured in articles in the New York Times, National Geographic, BBC, The New Yorker etc.
Areas of Research
- Artificial Intelligence
Understanding and designing structured representations: From LLMs to whales
The recent trend in NLP has been towards building general-purpose models; however, language (and other intelligent behavior) has a lot of structure—can we build tools that help us figure out how current models discover & represent this structure implicitly? Can we use what we’ve learned to make models better or refine our understanding of the world they describe? This talk will look at these questions from three perspectives. We will start by first understanding solutions language models intrinsically implement when trained on natural language data and how they thereby tie to the modelåÕs ability to generalize. Following that, we will look at how one can start to decipher the structure of (another black box language-like system), a naturally arisen communication system in animals, and, finally, use insights from these studies to equip embodied agents with a (latent) language of thoughtåÐhierarchical and compositional, and how it can enhance long-horizon reasoning and planning in these systems. _