Research

Publications

• Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs, Zhou Y., et. al. (under review)
• EinsteinPy: A Community Python Package for General Relativity, Astrophysics Source Code Library, Bapat S., et. al.
• Cross-Geography Generalization of Machine Learning Methods for Classification of Flooded Regions in Aerial Images, Arxiv, Sushant L. et. al.
• Music Emotion Recognition Using CNN-LSTM Architecture With Hybrid Spectral - Temporal Audio Feature Vector, IEEE Access, Pratyush K., et al. (under review)
• Multiclass Classification and Verification of Online Signatures, Research project under Dr. Ramya S., Presented as part of Industrial Training at Manipal Institute of Technology. (A+ Grade)

Research Projects

Active

• Robot Self Improvement via Large Language Models
  
  • Hypernetworks for Learning to Optimize: Developing new class of semantically augmented language model based optimizers for faster convergence in numerical and RL policy optimization tasks.
  • GenPlus - Code as Policy for Robotic Manipulation: Developing simulation environment to generate code as policy with Props based continuous feedback for precise control of robotic arm with movement primitive functions.
  • Co-developed Props and Props+ algorithms that combine numerical and linguistic reasoning for optimization of linear and non-linear reinforcement learning policies with 10 to 300 parameters across 10+ gym environments.

Past Projects

• Self Supervised Dense Point Tracking in Turbulent Videos
  

  • Implemented a test-time trained dense point tracking model robust to atmospheric turbulence by leveraging DINOv2 semantic features and RAFT optical flow model - leveraging the DINO-Tracker model - https://dino-tracker.github.io.

  • BMX Trees point tracking with atmospheric turbulence

• Music Emotion Recognition Using Hybrid Spectral - Temporal Feature Vector
  
  • Built a music emotion classification system based on Russell’s two-dimensional valence-arousal model using hybrid Convolutional Neural Network and Long Short Term Memory architectures.
  • Extracted a hybrid spectral-temporal feature vector involving audio fingerprints for training the model.
• Multiclass Classification and Verification of Online Signatures
  
  • Developed a robust signature verification tool that uses time-series data to detect forged signatures and classify signature owners using SVM.
  • Improved model efficiency by reducing the feature dimension using the Ramer-Douglas-Peucker sampling algorithm.