Utkarsh Uppal

I am a research intern at Nvidia where my research is focused on self-supervised learning for incremental task.

Previously, I worked as Machine Learning Researcher at Visio Lab, a Berlin based food checkout system startup and as a research intern under the guidance of Dr Mobarakol Islam and Prof Hongliang Ren, with research focused on model generalization and loss function optimization.

I recently graduated with Bachelors in Electrical and a minor specialization degree in Computer Science from Punjab Engineering College. Previously I,

• Spent 5 months [March-July, 2019] at Medical Mechatronics Lab, National University of Singapore, where I was advised by Dr Mobarakol Islam and Prof Hongliang Ren. During my internship, my research was centered around 3D Multi-Path Conditional GAN's and 3D Segmentation models for medical dataset.
• Pursued a 3 month [May-July, 2018] research internship at SSPL, Defence Research and Development Organization (DRDO), New Delhi.

Email  /  Google Scholar  /  LinkedIn

• [NEW]  Presented with the Institute Color Award
• [NEW]  TA for Fundamentals of Deep Learning workshop @ Nvidia GTC
• [NEW]  Started a research internship at Nvidia
• [NEW] Short paper on Normalised Label Distribution accepted at RSEML WS. @ AAAI 2021
• [NEW]  Volunteer @ NeurIPS 2020
• [NEW]  Awarded the IEEE Outstanding Student Volunteer Award
• MIT Hackathon Winning Team - Uboja got featured in True Africa
• Selected for KubeCon + CloudNativeCon Europe 2020 Scholarship
• Selected in the ExeCom of IEEE Delhi Section Young Professionals (IEEE YP AG)

My research interests lie at the intersection of Deep Learning, Computer Vision, and sustainable transition of AI to practical applications. Particularly, I am passionate about generalizable deep learning, self-supervised learning, curriculum learning, and sustainable reinforcement learning. I am also inclined towards application based DL & RL with focus on robotics and healthcare.

The vulnerability of models to data aberrations and adversarial attacks influences their ability to efficiently demarcate distinct class boundaries. We study the significance of changes in ground-truth distribution on the performance and generalizability of various state-of-the-art networks and demonstrate the role of label-smoothing regularization and normalization in yielding better generalizability and calibrated probability distribution.

We propose a 3D attention convolutional neural network (CNN) to segment brain tumor; extract novel radiomic features based on geometry, location, the shape of the segmented tumor, and combine them with clinical information to estimate the survival duration for each patient. Our proposed 3D attention module with decoder blocks consists of 3D spatial and channel attention in parallel to skip connections, thus exploiting the ubiquitous features of the receptive field.

Incorporating self-supervised learning for class incremental learning. The pre-text task training is based on random label augmentation and downstream task includes projection preservation along the orthogonal direction and feature enhancement by gaussian kernels.

The paper introduces a 3D conditional Generative Adversarial Networks (cGAN) to synthesize a 24-hour time-advanced Computed Tomography (CT) from the onset time CT of a patient with symptoms of Intracerebral Hemorrhage (ICH). We propose a novel 3D generator architecture fixated on learning both global and local features. Furthermore, we introduce an encoder-decoder type of discriminator framework to penalize the low confidence pixel-map more in comparison to high confidence with supplementary excitation module for improving the attributes of feature maps.

A Normalized Label Smoothing Focalloss (normLSF) to prevent overconfident learning and alleviate miscalibration in the genomic decision-making model. normLSF addresses the class imbalance issue and boosts the model generalization capacity. The proposed cost function is validated on two genomic datasets of the Chinese Glioma Atlas (CGGA) and the Cancer Genome Atlas Project (TCGA) and on CIFAR100 dataset.

Enhanced computer vision pipeline by integration of partial convolution based padding and smoothing in the network with UAV to monitor stubble burning.

Press T for Tetris