Machine Learning Researcher

Prithvi Tarale

I love drawing on biological principles to build efficient, elegant solutions for foundation models, lifelong learning, and beyond.

Building AI to
mimic Nature

I'm a fourth-year PhD student in Computer Science at the University of Massachusetts Amherst, advised by Dr. Sunghoon Ivan Lee. Before this, I studied CS at the University of Washington.

I believe nature solves problems in the most energy-efficient way possible, and that makes it the best place to find hypotheses for machine learning. Whether it's building bio-inspired foundation models for wearable sensors, designing decentralized multi-agent learning frameworks, or optimizing aircraft materials at Boeing, the through-line is the same: find how nature already solved it, then formalize that into a framework that works at scale. In my experience, the biological inspiration has meant smaller, faster, and more data-efficient ML systems than conventional approaches.

4.0 PhD GPA
2 Publications
1 Patent Filing
5+ Years in ML

Selected work

Bio-Inspired Self-Supervised Learning for Wrist-worn IMU Signals

Forty-Third International Conference on Machine Learning (ICML 2026)

Wrist motion is composed of elementary kinematic units called submovements — a principle from motor control theory. I used this to rethink how we tokenize IMU signals, treating movement segments as "words" for a Transformer to reason over. The result: a foundation model pretrained on 28,000 hours of wearable data from 11,000 participants that outperforms every existing self-supervised approach for human activity recognition.

+10–15% over SOTA 5× smaller model Bio-inspired tokenization
Read the paper

Distributed Multi-Agent Lifelong Learning (PEEPLL)

Transactions on Machine Learning Research (TMLR 2025)

What if AI agents could learn from each other instead of relying on expensive labeled data? I designed the first framework for decentralized multi-agent lifelong learning — agents that autonomously identify what they don't know, ask peers for help, and selectively learn from noisy responses. The framework includes a novel uncertainty quantification method using Variational Autoencoders that outperforms traditional entropy-based approaches.

17–20% over single-agent SOTA First distributed MALL framework Novel uncertainty via VAE
Read the paper

Industry · Patent

Neural Network Optimization for Aircraft Materials

Built an iterative neural network algorithm with Finite Element simulation in the loop to find optimal parameters for aircraft materials at Boeing Advanced Research Center. The work led to a patent filing by Boeing.

Project · 2023

Resilient Multi-Modal Learning with Missing Modalities

Designed a branched VAE and training regime to handle missing modalities during pre-training by reconstructing one modality from others. Improved baseline by 223%.

Where I've built things

Jan 2025 — Present

Research Assistant, PhD

Advanced Human & Health Analytics Lab · UMass Amherst

Building bio-inspired foundation models for wearable IMU data. Submitted Bio-PM to 2026.

Sep 2022 — Dec 2024

Research Assistant, PhD

BINDS Lab · UMass Amherst · GPA: 4.0/4.0

Designed the first framework for distributed multi-agent lifelong learning (PEEPLL). Published in TMLR 2025.

Oct 2020 — May 2022

Machine Learning Researcher

Boeing Advanced Research Center · Seattle, WA

Built an iterative neural network with Finite Element simulation to optimize aircraft material parameters. Work led to a patent filing by Boeing.

Oct 2020 — Oct 2021

AI Student Researcher

University of Washington · Seattle, WA

Enabled explainable AI customization through human-centric language. Trained a 90%-accurate few-shot classifier using SciBERT and RNN-attention.

May 2021 — Jul 2021

Research Fellow

Purdue University · Summer Undergraduate Research Fellowship

Predicted bugs in QT/OpenStack scripts. Measured fairness variance across identically trained models. Accepted to SURF Symposium.

Collaborations!

I love collaborating with and mentoring MS students. If you're interested in working on any of the areas below, I'd love to hear from you. Reach out and let's talk.

Foundation Models

Self-supervised representation learning, bio-inspired tokenization, and efficient pretraining strategies for time-series and sensor data.

ML for Health

Clinical applications, digital biomarkers, motor impairment detection, and wearable sensor analytics for healthcare diagnostics.

Lifelong Learning & Multi-Agent Systems

Distributed learning, uncertainty quantification, and designing frameworks for agents that learn collaboratively in dynamic environments.