Abhimanyu Dubey

I am a second-year graduate student in the Human Dynamics group at MIT, supervised by Professor Alex Pentland. My research interests are in robust machine learning and social cognition, and I also work occasionally in their applications in computer vision.

Prior to this, I received a master's degree in Computer Science and bachelor's degree in Electrical Engineering at IIT Delhi, where I was advised by Professor Sumeet Agarwal.

dubeya[at]mit.edu | github | twitter | scholar | linkedin


May 2019          Our paper on Bayesian bandits under heavy tail rewards was accepted to IJCAI 2019!
February 2019          Our work on adversarial defense via large-scale nearest neighbors was accepted to CVPR 2019 as an oral!
January 2019          Honored to receive the Snap Research Scholarship!
September 2018          Our work on regularization and fine-grained recognition was accepted to NeurIPS 2018!
July 2018          2 papers accepted to ECCV 2018!
May 2018          TuringBox has been accepted to FAT/ML 2018 as a full talk!


Thompson Sampling on Symmetric α-Stable Bandits

Abhimanyu Dubey and Alex Pentland
Real world data distributions arising from complex systems rarely are as well behaved as sub-Gaussian distributions. In this paper, we explore the Bayesian stochastic bandit problem in this context of badly-behaved 'heavy tailed' data. We develop efficient algorithms for Thompson Sampling in the K-armed bandit problem under symmetric α-stable reward distributions, and prove finite-time regret bounds of the same order as sub-Gaussian distributions.
International Joint Conference on Artificial Intelligence (IJCAI) 2019

Defense Against Adversarial Attacks using Web-Scale Nearest Neighbor Search

Abhimanyu Dubey, Laurens van der Maaten, Zeki Yalniz, Yixuan Li and Dhruv Mahajan
We study defense mechanisms that approximately project input images onto a hidden manifold by a nearest-neighbor search against a web-scale image database containing tens of billions of images to provide robustness against adversarial attacks.
IEEE Computer Vision and Pattern Recognition (CVPR) 2019 (Oral)

Maximum-Entropy Fine-Grained Classification

Abhimanyu Dubey, Otkrist Gupta, Ramesh Raskar and Nikhil Naik
We explore maximum-entropy discriminative training to reduce overfitting in fine-grained visual classification, where samples from different classes tend to be very similar.
Neural Information Processing Systems (NeurIPS) 2018
[proceedings] [arxiv]

Sparsely-Networked Evolution Strategies

Dhaval Adjodah, Dan Calacci, Abhimanyu Dubey, Peter Krafft, Esteban Moro and Alex Pentland
We modify the popular evolution strategies algorithm for deep reinforcement learning to introduce sparse network topologies of communication and demonstrate that sparse networks produce significantly higher average rewards as well as reduced communication costs.
NeurIPS 2018 Workshop on Deep Reinforcement Learning
[proceedings] [arxiv] [code]

Evaluating GANs on Explicitly Parameterized Distributions

Shayne O'Brien, Matt Groh, and Abhimanyu Dubey
We evaluate 16 different variants of Generative Adversarial Networks (GANs) on explicitly parameterized distributions in order to understand their similarities and differences.
NeurIPS 2018 Workshop on Critiquing and Correcting Trends in Machine Learning
[proceedings] [code]

Coreset-Based Neural Network Compression

Abhimanyu Dubey*, Moitreya Chatterjee* and Narendra Ahuja
We propose a method to compress neural network filters by recovering a sparse linear decomposition at the filter-level in CNNs, along with a data-dependent filter pruning algorithm to reduce memory usage and prediction run-time.
European Conference on Computer Vision (ECCV) 2018
[proceedings] [arxiv] [code]

Pairwise Confusion for Fine-Grained Visual Classification

Abhimanyu Dubey, Otkrist Gupta, Pei Guo, Ramesh Raskar, Ryan Farrell and Nikhil Naik
To reduce overfitting to sample-specific local artefacts in fine-grained visual classification, we propose a siamese-like architecture that attempts to bring inter-class logits closer to each other.
European Conference on Computer Vision (ECCV) 2018
[proceedings] [arxiv] [code]

Prediction Propagation for Domain Adaptation

Bjarke Felbo, Michiel Bakker, Abhimanyu Dubey, Sadhika Malladi, Alex Pentland and Iyad Rahwan
We tackle the problem of domain adaptation in natural language processing with imbalanced training data, using a label-based bootstrapping technique that attempts to learn large-scale domain differences by improving the correlation between predicted labels and domains.
ICML 2018 Workshop on Learning with Limited Labeled Data

TuringBox: An Experimental Platform for the Evaluation of AI Systems

Ziv Epstein, Blakeley Hoffman Payne, Judy Hanwen Shen, Casey Jisoo Hong, Bjarke Felbo, Abhimanyu Dubey, Matthew Groh, Nick Obradovich, Manuel Cebrian and Iyad Rahwan
We introduce a platform to democratize the study of artificial intelligence. On one side of the platform, contributors upload existing and novel algorithms to be studied scientifically by others. On the other side, examiners develop tasks to evaluate and characterize the outputs of algorithms.
International Joint Conference on Artificial Intelligence (IJCAI) 2018 (Demo)
Fairness, Accountability and Transparency in Machine Learning (FAT/ML) 2018 (Full Paper)
[proceedings] [website]

Sparse Matching for Embedding Image Macros

Abhimanyu Dubey, Esteban Moro, Manuel Cebrian and Iyad Rahwan
We create an algorithm to generate embeddings to capture the propagation and mutation of images on the internet, by decoupling underlying macros from overlaid adjustments and then combining decouple deep features. Our method opens up the possibility of obtaining the first large-scale understanding of the evolution and propagation of memetic imagery.
The International World Wide Web Conference (WWW) 2018
[proceedings] [arxiv]

Modeling Image Virality with Pairwise Spatial Transformer Networks

Abhimanyu Dubey and Sumeet Agarwal
To understand the regions of images that are responsible for promoting virality in images, we utilize pairwise spatial transformer networks that operate on a pyramid of multi-resolution images, and are hence able to capture image cues at different scales.
ACM Conference on Multimedia (ACM MM) 2017
[proceedings] [arxiv]

Deep Learning the City: Quantifying Global Urban Perception

Abhimanyu Dubey, Nikhil Naik, Devi Parikh, Ramesh Raskar and Cesar Hidalgo
We model urban perception of city scenes using a dataset of 1.3M pairwise comparisons over 6 perceptual attributes. We formulate the problem as a pairwise learning problem and use a siamese-like CNN architecture that learns to ''vote'' for images containing stronger attributes.
European Conference on Computer Vision (ECCV) 2016
[proceedings] [arxiv] [website] [code]

Examining Representational Similarity in ConvNets and the Primate Visual Cortex

Abhimanyu Dubey, Jayadeva and Sumeet Agarwal
We compare neural network representations with representations from the mammalian IT cortex and observe that improved performance on tasks like ImageNet corresponds to higher representational similarity to IT cortex.
ECCV 2016 Workshop on Biological and Artificial Vision (spotlight)

Deep Convolutional Networks for Modeling Image Virality

Abhimanyu Dubey and Sumeet Agarwal
We design a convolutional neural network architecture to model image virality on the internet.
ECCV 2016 Workshop on Web-scale Vision and Social Media

Coreset-Based Adaptive Tracking

Abhimanyu Dubey, Nikhil Naik, Dan Raviv, Rahul Sukthankar and Ramesh Raskar
We propose a constant-time method for real-time object tracking using coreset-tree representations of frames.
Preprint, 2015
[arxiv] [code]

(c) 2018 Abhimanyu Dubey , last updated February 25, 2019. Hits:web counter