Reshmi Ghosh

Logo

Senior Applied Scientist, Microsoft (MSAI)

   ABOUTACADEMICS   PROJECTS   Reviewing Activities


Papers

2025

  1. Bidirectional Human-AI Alignment: Emerging Challenges and Opportunities, Special Interest Groups (SIG) at Conference of Human Factors in Computing Systems (CHI) [https://dl.acm.org/doi/abs/10.1145/3706599.3716291]
  2. ICLR 2025 Workshop on Bidirectional Human-AI Alignment

2024

  1. Quantifying reliance on external information over parametric knowledge during Retrieval Augmented Generation (RAG) using mechanistic analysis - Blackbox NLP Workshop, Empirical Methods for Natural Language Processing Conference, 2024 [https://arxiv.org/abs/2410.00857]

Updates

</table>
[Personal] Attention based end-to-end speech to text model: Implemented the Listen, Attend and Spell paper to design a system for speech-to-text transcription using Locked Dropout, Teacher Forcing, and padded Cross-Entropy Loss.
[Personal] Machine and Deep Learning framework framework from scratch only using NumPy. </tr>
[DL research] Leveraged the VizWiz dataset (collection of 20,000 images captured by blind people through their mobile phones) to implement MobileNetv3 (CNN model) \& an LSTM based language model to develop a Visual Question Answering based framework. Achieved an accuracy of approximately 60\% using PyTorch library and NVIDIA K80 GPUs on AWS. </a>
[ML research] Conducted end-to-end data analysis to examine user behavior on Instagram about climate change by using unsupervised clustering techniques on web-scraped images \& associated hashtags. Determined that only ~10\% of all users posted relevant climate change related content between November-December'19.
[Personal-Kaggle] Developed an ML pipeline to conduct data-cleaning, pre-processing, feature engineering, \& built 'stacking' ensemble model using Logistic Regression, Support Vector, \& Random Forest classifier to predict survival rate in Titanic Disaster with ~93\% accuracy.
[Personal] The aim of the project was to augment enrollment strategy for Udacity. Coducted A/B Testing on a large dataset of students enrolling for *free* courses on Udacity to analyze the effect of 'studytime-filter' on students. Concluded based on statistically significant conversion rate numbers that filter shouldn't be launched.