Michael Cogswell

Michael Cogswell is a computer vision and machine learning researcher whose papers cover regularization, visual explanations and vision-language models. As a PhD student advised by Dhruv Batra at Georgia Tech, they were first author of the DeCov regularizer paper at ICLR 2016, second author of Grad-CAM at ICCV 2017, and a coauthor of Diverse Beam Search at AAAI 2018. Later papers include the DRESS vision-language model at CVPR 2024, a long-video retrieval study at WACV 2025 and a motion transferability benchmark at ICCV 2025. They are a research scientist at SRI International's Center for Vision Technologies in Princeton, New Jersey.

Now at: SRI International

Timeline

2015

PhD in Batra's lab

They join Dhruv Batra's Machine Learning and Perception Lab as a PhD student in 2015. Batra advises the PhD, and they work closely with Devi Parikh at Georgia Tech.
2016

The DeCov regularizer

As first author with Faruk Ahmed, Ross Girshick, Larry Zitnick and Dhruv Batra, they publish "Reducing Overfitting in Deep Networks by Decorrelating Representations" at ICLR 2016. The paper proposes DeCov, a regularizer that minimizes the cross-covariance of hidden activations to encourage non-redundant representations and reduce overfitting.
2017

Grad-CAM visual explanations

They are second author of Grad-CAM, published at ICCV 2017 with Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Devi Parikh and Dhruv Batra. The method uses the gradients of a target concept flowing into the final convolutional layer to produce a coarse map of the image regions that matter for the prediction, and it applies to image classification, captioning and visual question answering models.
2018

Diverse Beam Search at AAAI

They coauthor Diverse Beam Search, accepted at AAAI 2018, with Ashwin K Vijayakumar, Ramprasath R. Selvaraju, Qing Sun, Stefan Lee, David Crandall and Dhruv Batra. The method decodes a list of diverse outputs from neural sequence models by optimizing a diversity-augmented objective, and the paper reports results on image captioning, machine translation and visual question generation.
2024

The DRESS vision-language model

They coauthor DRESS, published at CVPR 2024 with Yangyi Chen, Karan Sikka, Heng Ji and Ajay Divakaran. The large vision-language model uses natural-language feedback from large language models, split into critique feedback for alignment with human preferences and refinement feedback for multi-turn interaction.
2025

Long-video retrieval and motion transfer

At WACV 2025, they coauthor "A Video is Worth 10000 Words" with Matthew Gwilliam, Meng Ye, Karan Sikka, Abhinav Shrivastava and Ajay Divakaran; its pipeline uses large language models to generate diverse synthetic captions for long videos, and fine-tuning on them gains 2.8% R@1 over the state of the art on ActivityNet with diverse captions. At ICCV 2025, they coauthor "Punching Bag vs. Punching Person: Motion Transferability in Videos" with Raiyaan Abdullah, Jared Claypoole, Ajay Divakaran and Yogesh Rawat, which evaluates 13 action recognition models on three motion transferability benchmarks.
As of October 2026

Research scientist at SRI

They are a research scientist at SRI International's Center for Vision Technologies in Princeton, New Jersey, where they work on deep learning for vision and language problems. They hold the same SRI Center for Vision Technology affiliation in 2025.