Tarun Dutt

I currently work on building robust deep learning models for medical imaging decision systems, extending work done during my Masters thesis. Some of my days are spent writing code to mine data from limb movements of fruit flies, working with Dr. Venkat Ramaswamy and Aman Aggarwal at the National Center for Biological Sciences (NCBS-TIFR).

I spent five wonderful years at the International Institute of Information Technology, Bangalore, graduating with a Bachelor's and Master's degree in Information Technology in July 2019, specializing in Data Science. I was advised by Prof. G.N.S Prasanna during my studies there.

Email  /  CV  /  Twitter  /  Google Scholar

profile photo

My interests are primarily in deep learning, vision, and computational neuroscience. I'm currently looking at techniques that allow decision systems to 'know when they don't know', taking me down the path of robustnes and generalization. I also dabble(d) in research involving quantifying locomotive behaviour in flies, Mixed Integer Programming for scheduling trains on the Indian Railway network and dynamical systems to model and study interactions between neurons.

Towards Artifact Rejection in Microscopic Urinalysis
Tarun Dutt, G.N.S. Prsanna, T.R. Dastidar, Ananth Shreekumar
Medical Imaging meets NeurIPS Workshop , NeurIPS, 2019

Developed and implemented robust deep convolutional neural net models to support the detection of unknown/unseen objects during prediction. Models were tested on a microscopic urine sample image dataset, where multiple artifacts present in the images were identified and rejected at test time, while also correctly classifying clinically significant objects. Techniques from the literature on open set recognition and out-of-distribution detection were benchmarked against a two stage localization mechanism to reduce misclasification of unseen samples, in addition to clustering experiments in feature space.

Markerless Limb Tracking in D. Melanogaster
Tarun Dutt, Venkat Ramaswamy, Aman Aggarwal

Conducting data-driven tracking and analysis of limb movements in Drosophila Melanogaster using image processing, clustering and time series techniques to quantify the development of locomotion in wild type and Parkinson's affected fruit flies (Drosophila Melanogaster). Developed heuristic algorithms to look at limb movement synchronizations, and combined spatial and temporal information to understand representations for inter-limb coordination.

Optimizing Train Schedules on a Large Railway Network
Sanat R, Tarun Dutt, Anushka C, Abhilasha A, G.N.S. Prasanna
21st International Conference on Intelligent Transportation Systems, 2018

Developed and rigorously tested a Mixed Integer Linear Programming optimization module to schedule new trains for the In- dian Railways, one of the largest and most dense railway networks in the world. Our model was able to double the throughput of trains moving over an important corridor in the network, while minimizing travel time and satisfying multiple crossover, overtaking and collision constraints.

A Faster Sampling Algorithm for Spherical k-means.
Rameshwar Pratap, Anup Deshmukh, Pratheeksha Nair, Tarun Dutt
10th Asian Conference on Machine Learning, 2018

Implemented and conducted experiments on multiple real world data sets, for a Markov chain based sampling algorithm that takes only one pass over the data, and gives close to optimal clustering similar to Spherical k-means++, i.e., a faster algorithm while maintaining almost the same approximation. The proposed algorithm is simple and easy to implement, and can be easily adopted in practice.

Analysis of a Coupled Neuron Model
Tarun Dutt, Balakrishnan Ashok

Built and simulated a coupled system of differential equations representing two types of biological neurons. Conducted detailed mathematical analysis on the behavioral stability, burst and spike patterns, and phase transitions of the dynamical system.

Augmented Reality Assistance in Healthcare
Tarun Dutt, TK Srikanth

Visualised head CT scan data using volumetric rendering with ray casting. Augmented the virtual 3D model onto video input featuring a patient in real time, to assist doctors during clinical procedures/diagnosis.

Graduate Student Instructor, Mathematics for Machine Learning, Fall 2018

Graduate Student Instructor, Learning and Cognitive Systems: An Optimization Perspective, Spring 2019

Website template from Jon Barron