Bayesian Reasoning And INtelligence (BRAIN)

Deep learning brought a lot of advancements in artificial intelligence, but have limitations in terms of learning from small data, handling uncertainty and model selection.  We develop Bayesian deep learning models which could overcome these limitations using Bayesian learning idea.   We  work on developing efficient inference techniques for Bayesian deep learning models.  On the applied side, we work on developing Bayesian deep learning models for problem arising in autonomous driving vehicles. 

We work on bringing Bayesian principles to Large Language Models (LLMs), allowing them to be more safe and robust by bringing uncertainty modeling capability in them. 

 Humans learn continuously over time, adapting and learning new tasks.  However, deep learning models exhibit catastrophic forgetting when trained on a sequence of tasks. We work on developing deep learning  models, which can overcome catastrophic forgetting through generative replay and  Bayesian learning techniques such as variational continual learning.  We work on continual learning problems in both vision and natural language processing. 

We also work on causal learning teachniques which allows the model to reason and become more robust.

 Bayesian non-parametric models  allow one to learn rich and flexible models due to their non-parametric nature and allow the model complexity to be determined by the data. This helps to overcome the problem of model selection to a great extent and helps to model uncertainty better.  Current  work focusses on developing  Bayesian non-parametric  models such as Gaussian processes and efficient inference algorithms  for them for various learning problems.  We develop  deep learning models such as deep Gaussian processes, and develop efficient inference algorithms for them. They can generalize from small data and can automatically determine model architecture to a great extent. 

With the widespread use of AI across various domains, we use machine learning and deep learning models  to solve various problems related to astrophysics such as Redshift prediction, and Galaxy morphology classification. 

We also work on poblems from mechanical engineering where we develop physics-informed neural networks for efficiently solving flow equations such as wave equations and Navier stokes equations. 

We can use machine learning to solve various real-world problems. One such problem is traffic data management. We use various machine learning and deep learning algorithms to understand and model traffic congestion and hence manage it efficiently.  We also work on analyzing and developing models for solving problems  arising in several other domains such as Telemetry data, ESG data,  and time-series data.