Arjun Mani

I am a PhD student in Computer Science at Columbia University, where I am co-advised by Carl Vondrick and Rich Zemel. I am interested in developing AI systems that can reason about the word in complex, human-like ways, and applying these systems to real-world problems in healthcare and climate change. My research is supported by the National Science Foundation Graduate Fellowship.

Previously, I completed my BSE in Computer Science at Princeton University with a minor in Applied Mathematics. I worked with Olga Russakovsky in the Princeton Visual AI Lab and Ryan Adams.

I am trained in Indian Classical Music - I perform vocal concerts in the US and India, and am passionate about spreading awareness of this art form. I also enjoy yoga, playing tennis, and hiking.


Point and Ask: Incorporating Pointing into Visual Question Answering

Arjun Mani , Will Hinthorn, Nobline Yoo, Olga Russakovsky

VQA Workshop 2021 (Poster Spotlight)


We extend Visual Question Answering (VQA) to questions requiring pointing, a key component of human communication, and introduce benchmark datasets and model designs for this new question space.

Geometric Algebra and Deep Learning: A Synthesis

Arjun Mani, Ryan Adams

Currently working on bringing the mathematics of geometric algebra into the computations and latent space of deep learning models. Our goal is to create more powerful and expressive internal representations using multivectors instead of vectors. Starting with word embedding models.

Adaptive Estimation of Active Contour Parameters Using Convolutional Neural Networks and Texture Analysis

Assaf Hoogi, Arjun Subramaniam*, Rishi Veerapaneni*, Daniel Rubin

IEEE Transactions on Medical Imaging, vol. 36, no. 3, March 2017

IEEE Homepage
Paper PDF

We improve on the level-set segmentation approach by using convolutional neural networks to adaptively modify parameters controlling for the expansion and contraction of the outline. The approach is fully automatic once an initial point has been provided. We achieve ~10% improvements on MRI and CT datasets of lesions.

CadML: Computational Antibody Design Using Deep Learning and Structural Protein Analysis

Arjun Mani, Stephen Meier, Thomas MacCarthy

Finalist, Regeneron Science Talent Search

We address the problem of predicting the effect of a mutation on an antibody’s binding affinity, in an attempt to identify beneficial mutations for antibody design. Our deep learning-based approach significantly outperforms existing methods.



Expressivity and Data-Dependency of Pruned Networks

ORF 569 Theory of Deep Learning, Fall 2021

Article coming soon!

Using random label experiments, we examine the extent to which pruning methods at initialization are data-dependent. We also examine the expressivity of these networks (e.g. their ability to fit true vs. random labels) and analyze their characteristics compared to networks pruned after training (e.g. using the lottery ticket hypothesis).

Majority Dynamics and Information Aggregation in Networks

COS 521 Advanced Algorithms, Fall 2020

Final Paper

We provide theory showing that certain graphs will never converge to a correct opinion even if each node is initially biased towards it; we also empirically examine majority dynamics and the effects of seeding or higher thresholds for changing opinions.

Examining Ambiguity in Human Pointing with Computer Vision

COS 529 Advanced Computer Vision, Spring 2019

Final Paper

I study how to understand pointing gestures that have ambiguous intent, examining how ambiguity can be explicitly predicted by semantic segmentation models.



AIDAN: Automated ML and Data Analysis with Voice Commands

Best Overall, HackPrinceton Spring 2018

Devpost Link

We built a chatbot that can respond to user voice/typed commands in realtime and perform linear regression, machine learning (SVM, logistic regression), and basic data analysis (mean/mode/etc.). Requires user only to upload a CSV file of the dataset.

DeepSquat: Deep Learning to Assess Exercise Technique

Best Health/Fitness Hack, HackPrinceton Fall 2017

Devpost Link

We built an app to assess squat technique, using a deep-learning based pose detection model.