projects | Ashish A. Singhal

Masters Projects

Augmenting Context-Aware Citation Recommendations With Citation and Co-authorship History.

Proposed a citation recommendation system utilizing co-authorship networks, employing Graph Convolution Networks for co-author and citation graph embeddings, and implementing transfer learning with BERT and Sci-BERT.

Suggesting Electric Car Charging Stations' Locations

Created a project with the goal of aiding government organizations and companies in identifying optimal locations for electric car charging stations by analyzing big-data datasets containing vehicle routes in coordinate format. Utilized the dataset to determine the most frequently traveled routes within the city, and strategically identified ideal charging station points based on user densities along these high-traffic routes.

Extreme Multi-Label Text Classification (XMTC)

Developed a hierarchical neural network structure to assign 20,000 Medical Subject Headings (MeSH) tree terms to biomedical papers, overcoming the issue of highly skewed data points for each label. Implemented a combination of Long Short-Term Memory (LSTM) and a fully connected neural network to predict labels at different levels within the MeSH tree.

Transfer Learning for Sleep Stage Detection

Addressing the challenge of limited patient data in a hospital setting, I investigated the application of transfer learning by fine-tuning machine learning models initially trained on publicly available datasets. Through experimentation, I identified the optimal layers within the model where parameter learning effectively takes place. Furthermore, I conducted a separate experiment involving the implementation of a Variation Autoencoder with KL-Divergence. This approach aimed to minimize discrepancies between the two datasets during classification tasks. Unfortunately, the unavailability of code stems from authorization issues.

Text Summarization

Developed an extractive text summarization model utilizing TF-IDF and cosine similarity, providing a hands-on learning experience in comprehensive text preprocessing techniques.

Personal Projects

Automatic Detection of Atrial Fibrillation (AF) Episodes

Conducted Time Series Analysis on ECG signals to extract features, addressing the challenge of an unbalanced dataset. Evaluated multiple models, providing a conclusive analysis on the effectiveness of certain models and identifying factors contributing to their superior performance.

Generative Adversarial Networks

Constructed a comprehensive and instructive blog tutorial, showcasing the implementation of a Generative Adversarial Network (GAN) to adeptly regenerate handwritten digits. This tutorial provides a hands-on and insightful exploration of GANs, specifically focusing on their application in digit regeneration.

Neural Machine Translation

Implemented a seq2seq model for German-to-English translation, incorporating an attention mechanism to handle longer text inputs effectively.

Deep Extreme Multi-Label Learning

Successfully replicated the state-of-the-art DXML paper for the thesis project, achieving 90% of the original score on the EurLex Dataset without access to the authors' hyperparameters, and submitted the code to PapersWithCode.com due to the unavailability of official code.