Representations & continual learning
Representation drift, feature alignment, and knowledge retention as models and data change.
MACHINE LEARNING · BIOLOGY · DRUG DISCOVERY
Learning representations
that remain useful.
I'm Anirban Pal, a PhD researcher at IIT Bombay Monash Research Academy working across machine learning, biology, and drug discovery. I study how models learn useful representations, retain knowledge, and generalize under distribution shift.
My research spans continual learning, self-supervised learning, and foundation models, alongside biological and molecular representation learning for cancer research, drug discovery, and catalysis.
Research directions
Foundation models & world models
Self-supervised features, latent prediction, and adaptation of pretrained representations.
Biological and Molecular Representation Learning
Learning from biological and chemical data for cancer research, drug discovery, and catalysis.
Selected work
All publications →Foundation-model representations
Studying what pretrained representations retain, how they change during adaptation, and how they transfer to downstream tasks. This work connects self-supervised learning and JEPA-based world models with continual learning and biological and molecular applications.
The Interface Contract: Repairable Geometry and Partially Predictable Drift in Continual Representation Learning
Pixel Reconstruction Keeps an Absolute-Coordinate Channel that Latent Prediction Lets Go
Action-Conditioned Sparsity for Latent-Space Planning
Valid Property-Enhanced Contrastive Learning for Targeted Optimization & Resampling for Novel Drug Design
A General Strategy to Access Pharmaceutically Relevant Pyridines from Nitroarenes
First online: 24 December 2025; issue date: 2 February 2026
Generating the Pd-catalyzed δ C–H chalcogenation of aliphatic picolinamides: systematically decreasing the bias
Recent work
- Two papers with Aarav Gupta on latent prediction and latent-space planning appear in the CAISc accepted-paper list.
- Work with Atharva Airen examines representation drift, test-time feature alignment, gradient reuse, and retrieval compatibility. Publication records →
- VECTOR+ introduces property-guided contrastive learning and latent-space sampling for molecular design.