Sai Ashish Somayajula

I am a PhD candidate in the ECE department at University of California, San Diego advised by Professor Pengtao Xie. The COVID-19 pandemic profoundly impacted me, inspiring my research journey. I realized that an efficiently trained language model for COVID-19 could rapidly analyze data and offer valuable insights for vaccine development, etc. Motivated by this idea, my research focuses on optimizing language models for healthcare applications. Training these models for healthcare involves unique challenges, such as limited labeled data, the need for models to adapt to the ever-evolving nature of diseases, and stringent privacy standards. My work aims to develop adaptable models that support clinical practice and medical research while ensuring compliance with regulatory requirements.

I obtained my Bachelor's degree in Electrical Engineering with a minor in Computer Science and Engineering from the Indian Institute of Technology, Hyderabad. I was fortunate to be advised by Professor Sumohana S. Channappayya and Professor Adity Siripuram. I secured the second-highest CGPA in the B.Tech program across all departments.

Email  /  CV  /  Bio  /  Google Scholar  /  Linkedin  /  Github

Master's students — shoot me an email if you want to discuss ideas or want mentorship, I would be happy to help!

News

• May 2024: Our work 'Token-Specific Watermarking with Enhanced Detectability and Semantic Coherence for Large Language Models' has been accepted to ICML'24!
• March 2024: Two papers accepted to NAACL'24!
• January 2024: Happy to share that our paper "Improving image classification of gastrointestinal endoscopy using curriculum self-supervised learning" has been accepted to the journal - Scientific Reports, Nature Portfolio.
• November 2023: Delighted to announce that I have successfully advanced to candidacy following my qualification exam. Committee: Pengtao Xie, Nuno Vasconcelos, Siavash Mirarab, Julian McAuley. My presentation slides can be found here.
• November 2023: My work on adaptive data augmentation for long COVID-19 literature classification has been accepted to the journal - Scientific Reports, Nature Portfolio. I would like to thank the editor and the reviewers for their insightful feedback on our work.
• November 2023: Gave a talk on my research at the IEEE YP Graduate Seminar Series - Session 4. The keynote speaker was Dr. Rangarajan Sampath, Senior VP and Head of the Center for Innovation in Diagnostics (CID) at Siemens Healthineers. Thank you for sharing your insightful thoughts on how to overcome failure. The video can be found here.

Professional Experience

• Apple Research Scientist Intern, 2023.
• Tencent AI Research Scientist Intern, 2022.

Research Direction

Watermarking in LLMs

With the proliferation of LLMs in generating synthetic datasets, distinguishing between human-curated and machine-generated texts is crucial to avoid misinformation. This distinction is particularly vital in healthcare applications, where the authenticity and reliability of data are of utmost importance. Statistically watermarking LLM-generated text can reliably detect such content. However, prior works face difficulty in achieving both high detectability and semantic quality of the generated texts after watermarking. We explore ways to improve both simultaneously.

A multi-objective optimization-based token-specific watermarking method to study and improve both watermark detectability and generation quality.

Dynamic Adaptability of Language Models

As diseases evolve and demographic data changes, language models must adapt to new data without losing accuracy on prior data. Traditional fine-tuning of LLMs on new, emerging data might lead to the model losing its prior knowledge of older data. We explore a research direction on how we can efficiently select a sub-network that, when fine-tuned on new data, will achieve maximum performance without significantly losing prior knowledge.

A bi-level optimization-based approach to finetune an automatically chosen sub-network within pre-trained language models on low-resource datasets to mitigate overfitting and reduce standard deviation. We are applying this method for continual learning in real-world medical application.

Parameter Efficient Fine-tuning (PEFT) methods

With the scaling of model parameters, such as transitioning from RoBERTa-large's 355 million parameters to GPT-3's staggering 175 billion parameters, fine-tuning becomes highly expensive in computation. PEFT methods become invaluable in such situations, within which LoRA is very effective. However, the LoRA method uses a predefined rank for each update matrix. We explore the research question of whether we can learn the optimal rank of these update matrices for a downstream task to improve performance.

AutoLoRA, a meta-learning framework designed to automatically determine the optimal rank for each low-rank adaptation (LoRA) matrix while parameter efficient finetuning.

Data-Augmentation

In healthcare, labeled training data is often limited—especially in cases involving emergent diseases, where timely and extensive data collection poses significant challenges. Data-augmentation methods can be invaluable in such scenarios. Within this, I explored two directions: 1) Can we leverage feedback from the downstream model to improve augmentation? 2) Can we synthesize the gradients of unseen words in the training dataset in a task-driven optimization without any external knowledge?

Introduce medical paraphrasing to augment data, coupled with a feedback mechanism. This approach utilizes a data-reweighting-based multi-level optimization framework with a meta-weight-network to enhance the classification performance of long COVID literature.

A bi-level optimization approach to synthesize gradients of unknown lexical information from known data, leveraging a task-dependent similarity matrix.

A data reweighting based domain adaptive feedback mechanism for end-to-end learning of text augmentation and classification models, overcoming traditional data augmentation limitations.