"Efficient Normalizing Flows: Theory, Algorithms, and Applications"

Dr. Sandeep Nagar
Institute for Advanced Study, Technical University of Munich

Wednesday, Feb 25, 2026

Colloquium -  499 DSL Seminar Room

03:30 to 04:30 PM Eastern Time (US and Canada)

Click Here to Join via Zoom
Meeting # 942 7359 5552

Zoom Meeting # 942 7359 5552

Abstract:

In this talk, I will discuss research that advances both the theoretical foundations and practical efficiency of Normalizing Flows, while also demonstrating their impact on computer vision tasks. We will discuss how to improve the efficiency and scalability of flow-based generative models. I will introduce a series of architectural and algorithmic, including mathematically grounded conditions for invertible 3×3 convolutional layers, a more expressive and efficient quad-coupling layer, and a fast parallel algorithm for general k×k inverse convolutions. Building on this, I will present an efficient backpropagation algorithm for the inverse of convolution, enabling a new training paradigm, Inverse-Flow, in which the inverse-of-convolution operation is used in the forward pass and convolution is used in the sampling pass. Together, these significantly reduce computational overhead for training and sampling while preserving exact invertibility and likelihood-based training.

Speaker Biography:

Sandeep Nagar is a Postdoctoral Researcher at the Institute for Advanced Study (IAS), Technical University of Munich (TUM). He received his Ph.D. from the Machine Learning Lab at the International Institute of Information Technology, Hyderabad (IIIT-H), where his research focused on probabilistic generative models, normalizing flows, and theoretical machine learning. His work centers on developing efficient, scalable algorithms for generative modeling with applications in computer vision.

Attachments:

File	Description	File size
sandeep_picture.jpg	Headshot	226 kB