Rahul Mohan Kumar

Logo


Rahul is a Data Scientist currently working as a researcher in the Social Neuroscience and Games (SNaG) Lab at the University of Colorado Boulder.

Skilled in all steps of machine learning engineering, Rahul enjoys building intelligent systems for better customer experiences and informed business decisions.

           


Research Projects


Falsifying the Causal Relevance of Linear Decoding Methods (Ongoing)

In this research project it has been discovered that the interaction between selectivity and the average activity of neurons provides a more accurate prediction of ablation performance deficits in groups of neurons across AlexNet, VGG16, MobileNetV2, and ResNet101.

Paper | Github


How are animate and inanimate categories differentiated in a pre-trained artificial neural network?

The primary objective of social neuroscience is to uncover neural mechanisms that underpin social interactions. Researchers have employed various analytical techniques to pinpoint brain regions responsive to social and nonsocial stimuli. Nonetheless, validating these findings necessitates investigating the impact of lesions on these identified brain areas. Unfortunately, such studies are impractical with human participants and not ideally suited to animal behavior. Therefore, this project addresses the validation of several analytical methods in identifying groups of neurons crucial to social functions by examining the effects of targeted lesions in an artificial neural network (ANN).

Github


Neural Network Ablation.

Detecting brain regions associated with particular cognitive functions, such as visual processing, can pose challenges. To address this, the project introduces modules designed for ablating pre-trained neural network models using PyTorch.

Github


Generate unit activations (Ex: Fairface dataset)

In this project, PyTorch is utilized to generate activations for all units within any layer of a deep learning model. As an illustration, the Fairface model is used, and the resulting activations are saved in a CSV file for further analysis.

Github


Personal Projects


Evaluate news articles that cover multiple perspectives of a topic

Effective knowledge discovery from newspapers, to comprehend the progression of specific events or identify analogous occurrences from the past, presents a significant and formidable challenge. This project focuses on enhancing topic tracking and clustering similar subjects to facilitate a comparative analysis of topic evolution among various groups. The outcome of this endeavor holds vital importance in opinion and trend analysis.

Github


SemEval-2022 Task 4: Patronizing and Condescending Language Detection

This project is my attempt on Task 4 at SemEval-2022, which was focused on detecting Patronizing and Condescending Language (PCL) towards vulnerable communities. In this project, I built a BERT sequence classifier and used PyTorch to fine-tune it.

Github


End to End speech synthesis system for Tamil

The Tamil Speech Synthesis system is built upon Google’s Tacotron model, utilizing keithito’s implementation available here.

Github