SUGANDHA SHARMA
  • Home
  • CV
  • Industry Experience
    • Microsoft Research
    • AMD
    • Nvidia
    • BlackBerry
    • Phoenix Interactive Design
    • Christie Digital Systems
  • Awards
  • News
  • Journal Club
I am a PhD candidate in MIT's department of Brain and Cognitive Sciences advised jointly by Prof. Ila Fiete and Prof. Josh Tenenbaum.  I am interested in exploring the computational and theoretical principles underlying higher level cognition and intelligence in the human brain, and leveraging them for building generalizable AI. 

Before joining the Fiete / Tenenbaum lab, I was at the Center for Theoretical Neuroscience, University of Waterloo (UW) where I completed my MASc with Dr. Chris Eliasmith, studying Theoretical Neuroscience and AI. Before that, I studied Electrical Engineering during my Undergrad at UW, with a minor in Management Engineering. 

I use mathematical tools to study how the brain helps us navigate the world. It’s fascinating that the same brain regions that help us navigate through a city, can also help us infer relationships in family trees and social hierarchies. The brain continuously computes the body’s position in space and makes adjustments to that estimate as we move about.  I am particularly interested in how the brain extrapolates information from one spatial environment to navigate new and different environments.

Our ability to navigate a labyrinth, for example, depends on a so-called “cognitive map,” or a mental representation of our physical environment. I study how this map is learned and organized in the brain so that we can quickly and efficiently find our way in the physical — and social — world. My goal is to build AI that can do the same with equal or greater efficiency than humans.

I am currently exploring the coding principles in the hippocampal circuits implicated in spatial navigation, and their role in cognitive computations like structure learning and relational reasoning useful for SLAM (simultaneous localization and mapping) but also in more abstract relational domains. 
CV.pdf
File Size: 116 kb
File Type: pdf
Download File

Highlighted Research

Picture
Human-agent alignment in large-scale multi-player games. 
​(@Microsoft Research, in preparation)

​

​GPT based g​ame-playing AI agents for a 4 x 4 multiplayer game called Bleeding Edge.

​
Picture
​GPT (Generative Pre-trained Transformer) based AI architecture

  • Harnessed unsupervised manifold learning on extensive game-play data to analyze behavioral alignment between humans and AI agents - over 100,000 games.
  • Leveraging this insight, LLM based AI agents were trained using imitation learning, for targeted behavior replication.
  • Throughout, GPT-4 was used for hypothesis formulation, validated against human and AI datasets.

Picture
Map Induction: Compositional spatial submap learning for efficient exploration in novel environments
​(ICLR 2022)

Efficient exploration by a Map Induction based AI agent
Picture


AI agent uses generative inference and approximate belief space planning in a model-based RL setup to induce unobserved regions from prior experiences.

The planner approximately solves an underlying POMDP (partially observed markov decision process) using the MCTS (monte carlo tree search) algorithm, and significantly improves the exploration performance of the state of the art POMCP (partially observed monte carlo process) based planning models. 




​Cognitive science experiment to test the Map Induction hypothesis in human subjects exploring 3D environments.

Picture
Content addressable memory without catastrophic forgetting by heteroassociation with a fixed scaffold.
(ICML 2022)


MESH can be used for memory retrieval in LLMs (Large Language Models)


Architecture is akin to an autoencoder with constrained activations and one-shot learned weights. 


Picture




​MESH exhibits a smooth tradeoff between number of memories and their richness.

Picture
Picture
Picture
Picture
Picture
Model of hippocampal episodic memory unified with and enabled by pre-structured spatial representations.
(@MIT, in preparation)



​





​AI agent that uses modular grid cell encoding to learn long overlapping sequences.





​



​

​AI agent that uses grid and place cell codes to learn spatial maps of environments as composed of map primitives (fragments).


​

The agents learns primitives that can be used compositionally to represent new environments, thus leading to generalizable learning of spatial maps through unsupervised one-shot learning.
Proudly powered by Weebly
  • Home
  • CV
  • Industry Experience
    • Microsoft Research
    • AMD
    • Nvidia
    • BlackBerry
    • Phoenix Interactive Design
    • Christie Digital Systems
  • Awards
  • News
  • Journal Club