My research

Throughout my career, I have been fortunate to contribute to various research directions, including machine translation, natural language processing, imitation learning, reinforcement learning, optimization, offline reinforcement learning, and machine learning infrastructure and systems. A detailed list of my publications can be found on my Google Scholar page.

Currently, my main research interests are in building scalable machine-learning systems that are aligned with a feedback. The feedback might be coming from either humans or as a result of some form of algorithmic computation. On this page, I also list three pillars of my research that I believe will be crucial to address as we integrate AI algorithms into the real world and societies.

The three research pillars

Efficiency

Achieving human-like sample efficiency while significantly improving compute efficiency is essential to reduce AI’s environmental impact and ensure its scalability and sustainability for future advancements.

Robustness and safety

Robustness and safety are critical for AI algorithms as they transition into real-world applications, ensuring reliable performance and minimizing unintended consequences that could profoundly impact society and the environment.

Reasoning and Discovery

System-2 level reasoning abilities in AI are crucial for accelerating scientific discovery, enabling the generation of deep insights, hypothesis testing, and complex problem-solving akin to human analytical thinking. Reinforcement learning is core for this research pillar.

My Research Vision

Build safe, robust🦾 agents🤖 that can make use of the experiential📝 data efficiently 🏎 to make a positive impact.

Efficiency

Achieving compute and human-like sample efficiency via:

Better architectures
Hardware-aware algorithms
Improved learning paradigms and losses
Better test-time decoding algorithms