Tanushree Banerjee

Selected Independent Work Reports

Earlier during my undergraduate studies at Princeton I worked on bias in visual understanding pipelines under Prof. Olga Russakovsky in the Princeton Visual AI Lab.

This paper proposes several approaches to identify questions unrelated to an image to prevent object hallucination in VQA models. The best approach achieved a 40% improvement over the random baseline.

This paper analyzes the bias in a model against underrepresented skin tones in the training dataset on the skin lesion classification task.

Selected Course Projects

These include unpublished research-related work as part of semester-long course projects during my undergraduate studies at Princeton.

We reproduce the results of “METRA: Scalable Unsupervised RL with Metric-Aware Abstraction” (Park et al.,2024), which proposes a novel unsupervised RL objective, called Metric-Aware Abstraction (METRA) (Park et al., 2024), which learns diverse, useful behaviors, as well as a compact latent space that can be used to solve various downstream tasks in a zero-shot manner. We validate the claims of this paper in our reproduction study, along with additional figures to confirm their claims.

We conduct experiments to understand the effect of errors in speaker diarization and speech recognition on an LLM’s summarization performance. We conduct counterfactual analysis by automatically injecting speaker diarization or speech recognition errors into spoken dialogue.

We propose and evaluate two dataset and model-agnostic frame sampling strategies for computationally efficient video action recognition: one based on the norm of the optical flow of frames and the other based on the number of objects in frames.

We empirically analyze what aspects of the in-context demonstrations contribute to improvements in downstream task performance, extending the work of Min et al., 2022 to multiple choice and classification tasks.

We reproduce the results of “Double-Hard Debias: Tailoring Word Embeddings for Gender Bias” (Wang et al.,2020) to reduce the gender bias present in pre-trained word embeddings. Additionally, we evaluate the proposed technique on Spanish GloVe embeddings to assess whether these debiasing methods generalize to non-English languages.

Teaching & Outreach

I have been incredibly fortunate to receive tremendous support from faculty mentors while at Princeton to pursue a research career in computer science. I am keen to use my position to help empower others to pursue careers in computer science. This section includes my teaching and outreach endeavors to help address the acute diversity crisis in this field.

As a UCA, I held office hours for students in the seminar, helping them debug their code and advising them on their semester-long independent work projects.

As an instructor, I developed workshops and Colab tutorials leading up to an NLP-based capstone project for this ML summer camp for high school students from underrepresented backgrounds, focusing on the potential for harm perpetuated by large language models. I also organized guest speaker talks to expose students to diverse applications of ML in non-traditional fields.

Design and source code from Leonid Keselman's Jekyll fork of Jon Barron's website.
Updated June 2024.