Focus Period lund 2026
Postdoctoral Researcher
Technical University Munich (Germany)
Manish Krishan Lal is a postdoctoral researcher at the Technical University of Munich, working with Suvrit Sra. His research focuses on the mathematical foundations of optimization and machine learning, with an emphasis on projection-based methods and nonconvex geometry arising in modern learning paradigms. He received his PhD from the University of British Columbia, where he worked with Heinz H. Bauschke and Xianfu Wang.
Presenting: Projection-based framework for learning, inference, and sampling in neural systems
For training, we develop a method for neural networks on Boolean data and continuous pedagogical datasets. Instead of optimizing a loss function, the approach formulates training as a nonconvex constraint satisfaction problem. We compare this method with state-of-the-art gradient-based baselines and discuss both its advantages and limitations.
For inference in discrete systems, we demonstrate the approach across several tasks, including multiplier-circuit discovery, binary autoencoding, logic-network inference, and cellularautomata learning. In these settings, the method achieves exact solutions or strong generalization in regimes where standard gradient-based methods struggle. These results suggest thatprojection-based constraint satisfaction offers a viable and conceptually distinct foundation for learning in discrete neural systems, with potential benefits for interpretability and efficientinference.
Finally, we present new projection-based algorithms for sampling, with illustrative examples in diffusion models.