Ritwik Bera

I am a graduate researcher at the Vehicle Systems and Control Laboratory at Texas A&M University. I work on human-in-the-loop machine learning in close collaboration with the Human Research & Engineering Directorate at Army Research Lab, Aberdeen, MD. During Summer 2020, I worked on leveraging human input modalities such as eye-gaze to make robots more capable at understanding intent. I have also been involved in developing generative modelling-inspired architectures that help robots learn composable skills from unstructured demonstrations (Read more here).

An eagerness to dive deeper into the fundamental data structures and algorithms that enable scalable machine learning led me to working on a repository of implementations of such concepts. These range from scripts that explore usage of data structures like KD Trees in search applications to snippets illustrating probabilistic algorithms for sampling, shuffling etc. Take a look here and here.

In an attempt to reduce boilerplate code across projects, I have also invested some time in building PyTorch-based tools that I found to have utility in multiple ML projects. These range from memory profilers to computation graph visualizers. All of this is available here.

A ‘whatever it takes’ approach in ML projects has led to me being a ‘full stack’ ML person. My work has encompassed building scalable data pipelines, conceptualizing novel model architectures, building custom visualization tools and engineering work for predictable, fault-tolerant deployment of proof-of-concept solutions.

My academic training coupled with quite a bit of self-learning has provided for a ‘first principles’ approach to machine learning, one that is rooted in information theory, data structures and statistical analysis.

I occasionally like to spend my time reading about what’s new in applied research; how some core concepts in deep learning are making an impact in real-world products. This could range from understanding how Graph Neural Networks could be used for decision-making in self driving cars and improving route predictions in Google Maps to even getting my hands dirty by working on ‘from-scratch’ implementations of concepts (like this implementation of RingReduce used in Baidu’s DeepSpeech work).

Outside working on projects and writing code, I also present and write on topics within the machine learning subfield.

My CV can be found here.