Projects

Developed a novel method, EmbeddingCAM, for visualizing embedding networks, significantly enhancing understandability and trust in model predictions. The approach involves defining a custom loss function and using proxies as class representatives, enabling heatmap generation. Effectiveness was demonstrated on the CUB-200-2011 dataset, where it achieved competitive results and highlighted relevant image regions. This method helps explain embedding network predictions, advancing explainable AI.

Generated state-of-the-art model for Visual Question Answering by training an attention based architecture. Built end-to-end pipeline for preprocessing, training and validation in Tensorflow. Optimised the model with custom divergence loss and multiple regularisation techniques like $latex L_2$ Normalisation, Dropout and Weight Clipping. Achieved 58% accuracy on VQA 2.0 dataset.

Developed a machine learning system to predict the existence of exoplanets, using dips in light signals of the stars. Extracted features using Wavelet Transform and achieved the F1-score of 0.96 with AUPRC score of 0.97. Developed an outlier removal algorithm and used Temporal Convolutional Neural Networks to predict exoplanets. Won bronze medal for the solution in Inter IIT Tech Meet 2018

Formulated a multi-period fuzzy portfolio to minimize risk and maximize wealth with minimum transaction costs. Used investors’ aspiration levels to resolve objectives and used Genetic Algorithm to solve multi-objective non-linear integer programming. Implemented the model on Indian Stock Market and generated optimum frontier using NSGA-II

Developed a decentralized location based image sharing platform using Ethereum and IPFS. Designed smart contracts using Solidity for storing and processing information on Ethereum network. Used IPFS (InterPlanetary File System) for storing the image in decentralized network. Created web interface using embark.js to browse and upload images.

Applied multi-level stacked generalisation on machine learning models to predict relevance score of a search query. Ranked in top 6% participants and received bronze medal in the competition hosted on Kaggle. Created complex pipelines for generation of more than 200 NLP features like TF-IDF, Counting Features, Distance Features, etc. Used hyperopt for tuning hyperparameters of complex models like extreme gradient boosting and neural network

Integrated tweets with news to create an engaging news browsing platform. Used hybrid semantic similarity based on LSA and WordNet to rank the tweets correlated to news. Devised an iterative algorithm to extract hidden topics from Twitter. Designed interactive web-based dashboard to provide temporal and spatial analytics.