Shubham Anilkumar Jain

Developed a Python library with several state-of-the-art explainability methods that can plugin into any deep learning image classifier and would show the parts of the image that influence the decision. This provided an insight into what the models were learning.

The library was executed to be easily installed and usable with any PyTorch deep learning classifier. It included more than four deep learning methods for explainability developed based on the published research.

Developed a large-scale privacy-preserving analytics platform with a small team located worldwide. The platform allows authorized users to use telecom data to generate meaningful insights like population density in several regions. Policymakers then used these insights in those countries.

The platform was deployed in Senegal and Colombia with our telecom partners.

We devised a new methodology to evaluate the robustness of perceptual hashing algorithms used to detect similar images. We showed that an attacker could modify the image such that it visually looks similar to the original image but cannot be detected by the matching perceptual hashing algorithm. Our methodology and results have been published at a top-ranked security conference.

A project to evaluate several evolutionary strategies, such as particle-swarm optimization and covariance matrix adaptation-based strategies. We built an environment to simulate 2-player Othello agents, similar to an Open-AI gym. This simulator was then used to train and evaluate several different agents.

Design an NLP algorithm to extract rules from a language to convert it to UNL (Universal Natural Language) format. UNL is often seen as the HTML of future linguistics. One way to do machine translation would be to encode a language to a UNL format and decode UNL to any other language. Based on this idea, we built a rule-based converter for Hindi to UNL. We then used search to extract cases where rules failed and then used them to increase the rules in the dataset, thus creating a more comprehensive set of rules that allowed us to create machine translation models when there are not enough datasets for deep learning models.

We used tools like NLTK, Stanford Parser, Tokenizers, and others to achieve our goals.