Hi, I'm Ansh! I'm an undergraduate student at UIUC majoring in Computer Science. My research interests include meta-learning, computer vision, medical imaging, ML for healthcare, and ML for drug discovery. I am currently advised by Professor Yuxiong Wang and Professor Shenlong Wang, and am working as an ML Intern at AbbVie.

I tentatively plan to graduate with a B.S. in Computer Science in Spring 2024 and am currently applying for MS/PhD programs that start in Fall 2024. I am also on the lookout for research internships for Summer 2024. If you're interested in working with me, please reach out!

E-Mail | Github | LinkedIn | Resume

YouTubePD: A Multimodal Benchmark for Parkinson’s Disease Analysis
Andy Zhou*, Samuel Li*, Pranav Sriram*, Xiang Li*, Jiahua Dong*, Ansh Sharma, Yuanyi Zhong, Shirui Luo, Maria Jaromin, Volodymyr Kindratenko, Joerg Heintz, Christopher Zallek, Yuxiong Wang
Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track. (NeurIPS 2023)
paper
website

Deep-Learning Enabled Assessment of Neurocognitive Performance in Object Following in Mixed Reality
Ansh Sharma*, Keerthana Nallamotu*, Mukhilshankar Umashankar, Shenlong Wang, Inki Kim
The Seventh IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies. (CHASE 2022)
paper

AI based imaging biomarker development in ADPKD mouse model
Snehal Vadvalkar, Ansh Sharma, Xiaomeng Zhang, Kaneli Galiotos, Bradley Hooker, Dustin Wooten, Varsha Mohan, Quaisar Ali, Ana Basso, Eric Mohler, Abhishek Pandey
World Molecular Imaging Congress (WMIC 2022)
abstract

Software Engineer Intern @ Amazon Translate

• Evaluated and fine-tuned various large language models (LLMs) for automated translation quality evaluation tasks.
• Reimplemented research papers and worked with applied scientists to evaluate techniques on internal LLMs and datasets.
• Created an LLM-based scoring workflow in the translation model deployment pipeline to compute automated quality metrics.

Machine Learning Intern @ AbbVie

• Researched representation learning models for cell painting using self supervised contrastive learning and arcface loss.
• Implemented and trained a ResNet based architecture on terabytes of molecular perturbation data using PyTorch and AWS.
• Transfer learned a 3D U-NET model to segment kidneys and cysts from MRI scans for tracking PKD progression in mice.
• Utilized multiomic data (MRI scans, RNA-Seq, blood/urine biomarkers) to cluster and predict response to drug treatment.
• Identified potential biological pathways for a novel treatment via gene ontology analysis, enabling further drug optimization.

Visiting Student @ Argonne National Laboratory

• Containerized and deployed an AlphaFold2 Docker image on the Delta Supercomputer for use by biology researchers.
• Improved inference pipeline runtime from 75 minutes to 20 minutes through CPU parallelization and multiple GPUs.

Software Engineer Intern @ AWS AppFabric

• Developed and deployed a full-stack prototype for an upcoming greenfield AWS product using Java, TypeScript, and React.
• Implemented internal authentication protocols and multi-origin CORS handling to allow call access to a Lambda API.
• Designed and integrated an interactive front end into an existing AWS product to demonstrate functionality for stakeholders.

Neural Music Transcription with Spatiotemporal Vision Models

• Designed and implemented U-Net, CRNN, and Transformer based architectures to transcribe piano audio files into notes.
• Evaluated qualitative/quantitative performance of each model with differently weighted losses to reduce data imbalance.
• Built data processing pipelines to convert 200 Hours of audio data into mel-spectrograms to train and evaluate models.

Meta-Learning For Regression via Data Re-Weighting

• Designed a data-reweighting algorithm to alleviate test-train imbalance for regression tasks using meta-learning techniques.
• Tested algorithm on an age-detection task (UTKFace) for proof of concept and outperformed proposed baselines.

Style Share

• A website that allows users to generate 3D scenes and stylize them according to the style of another image using ML.
• Used quantization and distillation to reduce TF.js model size and improve speed by 4x while running within browsers.
• Included authentication with Google OAuth and a gallery to upload and share photos using a cloud storage bucket.

Infected and Detected

• HackIllinois 2022: Best Community & Sustainability Track Project.
• An ML based edge computing tool to help farmers get analytics on their crop health and invasive weed growth over time.
• Trained an image classifier using transfer learning on MobileNetV2 to identify different types of plant diseases and weeds.
• Used pruning and quantization to shrink model size further by 10x to run quickly on a Rasberry Pi with a Coral Edge TPU

Template credits: Jon Barron