Harshveer Singh Email & Phone Number

Bengaluru, KA, IN

Bengaluru, Karnataka, India

Gurugram, Haryana, India

Gurugram, Haryana, India

Worked on building a quasi-linear attention model to isolate 'interesting' events from the background during the collision of protons with low-Z targets.

Adversarial training for Facebook's blender botAiming at a self play regime for conversational agents, and extending that to competitiveconversation where an agent discriminates the output distribution of the other agentagainst human dialogue distribution.

Poly encoder regime for fine-tuning decoder-only model (GPT-2)Decoder model fine tuned like such, apparently is more robust to inductive bias thanencoder model even though encoder reached better recall@k/C score.

Patiala, Punjab, India

Planned the academic curriculum: topics, number of classes, mentor assignment.Mentored the Machine Learning research groupMentored teams for PECFEST Hackathon ‘19, SATURNALIA Hackathon ‘19

Mandi, Himachal Pradesh, India

Project - Analytical study of success of Batch Norm

Patiala, Punjab, India

Developed TiNotifications, an event management and publicity platform for Thapar Institute of Engineeringand Technology, (currently in use, exclusive for TIET students), backboned with aTransformer based attention model (Poly Encoders), to pick the next favorable task to operateon, ensuring the smooth functioning of a given student organization at TIET.

Analytical study of success of Batch NormShowed that batch normalization smoothes the loss surface and how it brings that effect,through the study of eigen-values of the hessian of input matrix.

Developed a novel regime for varying beta2 hyperparameter of Adam (adaptive momentum), preventing Adam from getting stuck in sub-optimal minima. A similar result was also shown in the parallel but completely independent research Sashank J. Reddi, Satyen Kale & Sanjiv Kumar (ON THE CONVERGENCE OF ADAM AND BEYOND) theorem 3.

Ludhiana, Punjab, India

Design, Develop, maintain, and improve internally and externally facing web based products with moderate guidance and direction from manager and peersTools used through the time period: NodeJS | ReactJS| VanillaJS | MongoDB | Nginx | Multiple Third-party APIs