A Non‐Chaotic Pruning Strategy for Neural Networks

Pruning strategies such as L0, L1, Rank based pruning change the feature importance orderings of Neural Networks in chaotic manners. This project hypothesizes that Granger‐Causality based pruning may be a non‐chaotic neural network pruning strategy.
Supervised by: APPCAIR - BITS Pilani

Assessing the Aptitude of Language Models in Comprehending Advertisements

[ongoing] Advertisement media are fundamentally different from typical videos and images. They are more than just their content, persuade users to take certain actions, and often use creative atypicalities to deliver their message.
Advertisement images from the Kovashka Ads Dataset were textually verbalized. These were presented to text‐based language models like GPT‐3.5, GPT‐4, and FLAN‐T5, which were evaluated on Action‐Reason pair and Atypicality understanding tasks.
As a comparison between text and vision, Vision‐Language models like BLIP2 were also evaluated on the advertisement images.
Supervised by: Adriana Kovashka - University of Pittsburgh

Semi‐Supervised Segmentation and VQA on Aerial Flood Images

[ongoing] Designed a semi‐supervised Image Segmentation and Graph based Visual Question Answering system for the FloodNet challenge.
The Segmentation network is based on CutMix and Cross Pseudo Supervision.
The VQA system applies geodesic dilation and morphological operations on the segmentation maps. Connected component counts are per‐formed on 4‐adjacency graphs made from the processed segmentation maps.
Supervised by: Sravan Danda - BITS Pilani