Harshvardhan

MS Data Science

EPFL

Biography

Harshvardhan is a Master’s in Data Science Student at EPFL, working on optimization methods in machine learning.

Interests

Machine Learning
Learning Theory
Optimization

Education

MS in Data Science, 2021

Ecole Polytechnique Federale de Lausanne (EPFL)
BTech in Electrical Engineering, Minor in AI, 2019

Indian Institute of Technology, Kanpur (IIT Kanpur)

Publications

Srujan T Thomdapu, Harshvardhan, Ketan Rajawat

December 2020 arXiv

Stochastic Compositional Gradient Descent under Compositional Constraints

Vikas Singh, Harshvardhan, Nishchal Verma, Yan Cui

July 2018 In FUZZ-IEEE 2018

Optimal Feature Selection using Fuzzy Combination of Feature Subset for Transcriptome Data

Projects

Constrained Stochastic Compositional Gradient Descent

Convergence analysis of constrained formulation of compositional SGD

Optimal Batch size per iteration

Obtained a theoretically-optimal batch size per iteration for finite-sum mini-batch SGD for convex objectives. Analysed optimal per iteration v/s same optimal batch size for all iterations, establishing a theoretically optimal batch size framework for SGD with batch size increasing with error.Established batch size and step size equivalence for convergence-diagnostic based adaptive step size and batch size control algorithms for SGD. Designed a new gradient norm based test for detecting SGD convergence. Tested batch size control and GradNorm test to obtain comparable performances.

Robust Training for Neural Networks

Alt-opt based 1-layer ReLU training with LS for +ve and -ve weights and sparse recovery to choose the weights' signs.

Adversarial attacks on qSGD and EF-signSGD with gossip communication

Implemented qSGD and EF-signSGD on MNIST for distributed gossip based optimization with byzantine nodes which send reversed gradients, and median-based and fractional mean based protection schemes for ring and torus.

Junior Machine Learning Engineer at Visium SA.

Developed the API, and machine learning models for serving document extraction from contracts using BERT and time series forecasting using CNN and tsfresh based architectures.

Stochastic Hamiltonian Descent

Proved convergence of a stochastic variant of the First Explicit Method of Discretization of Hamiltonian Descent Methods with same bounds as the deterministic version under certain strict stochastic assumptions.

MyFoodRepo

The MyFoodRepo project aims to produce accurate nutritional value estimates from images of food items. My work involved training the baseline Tortilla image classifier on a private dataset of food images, handling class imbalance in the data, setting up a coco-ui like segmentation interface on MTurk, and a webcrawler for obtaining further training data.

Direction of Arrival and Source Position Estimation

Implemented papers for estimation of direction of arrival and position of source in a room from microphone array.

Outlier Detection Using Loggarma Models

Implemented outlier detection by fitting time series to a LogGARMA model.

Recommend a Reviewer

Implemented an alt-opt based recommendation system to match authors to papers based on NeuRIPS data.

Feature Selection using fuzzy combination of fuzzy subsets

AUV

Designed Battery Management System for Autonomous Underwater vehicle team at IIT Kanpur.