Research projects undertaken

Academic research I've conducted in Data Science, Computational Social Science, AI, ML & NLP, listed in reverse chronological order.

Visiting Researcher | Nokia Bell Labs (Social Dynamics Group), Cambridge

Bell Labs Logo July 2023 – Present

Project 2: Studying health discussions on social media

July 2023 - Present
Tech Stack: Python, NumPy, Scipy, Pandas, Matplotlib, Seaborn

  • Ideating & implementing ways to study various forms of engagement in health discussions on social media

health_project_img

Project 1: Unpacking dream content on Reddit (Link)

March 22 - April 23
Tech Stack: Python, NumPy, Pandas, Spacy, NLTK, NetworkX, Matplotlib, Plotly, Gephi

  • Developed a methodology using BERTopic to extract common topics & themes in self-reported dreams on Reddit. (Research accepted & has been presented at IC2S2 conference, Copenhagen)
  • Studied the prevalence of extracted themes across time & across vivid, recurring, lucid dreams & nightmares
  • Created co-occurrence networks of common themes & health symptoms in dream reports to visualize 40k dreams

subreddit_colab_pipeline

Machine Learning Research Intern | Jadavpur University, India

Jadavpur University Logo September 2020 - April 2021 & January 2022 - March 2022

Project 3: Emotion detection from EEG signals using hybrid 1D-CNN + LSTM

February 2022 - March 2022
Tech Stack: Python, NumPy, Keras, Scikit-learn, Matplotlib

  • Subject Dependent Approach; Accuracy: 94.11% for valence, 94.39% for arousal (emotion dimensions)
  • Subject Independent Approach; Accuracy: 90.09% for valence, 91.39% for arousal

eeg_1d_cnn_lstm_pipeline

Project 2: Automatic Vehicle Detection using feature ensemble of pre-trained networks

January 2022
Tech Stack: Python, NumPy, PyTorch, Scikit-learn, Scikit-image, OpenCV, Matplotlib

  • Classification of patches of images of road scenes in structured (GTI) & unstructured driving scenes (IDD)
  • Method 1: Light-weight feature-extractor - Histogram of Oriented Gradients (HoG) with Linear SVM - 89.73% accuracy
  • Method 2: Ensemble of 3 pre-trained CNNs as feature extractors with Linear SVM - 99.83% accuracy

dfe_avd_pipeline

Project 1: Comparative Analysis of Generative Adversarial Networks for Bangla Character Generation

September 2020 - April 2021
Tech Stack: Python, NumPy, PyTorch, Scikit-learn, Matplotlib

  • Trained 3 GANs : DCGAN, ACGAN & WGAN-GP for Bangla character generation
  • Dataset used : BanglaLekha-Isolated (84 bangla characters : Numerals, basic & compound characters)
  • Performance comparison using Inception Score (IS) & Fréchet Inception Distance (FID). Fine-tuned Inception Model prior to evaluation for better evaluation.
  • Preliminary attempts have been made to explore the latent space using vector arithmetic to generate unseen compound characters. Some samples for the same, can be seen in the following figure.

bangla_dcgan_vector_arithmetic