Gautam Machiraju

Research Themes

My research interests broadly span AI-human collaboration and AI-driven knowledge discovery to expand our understanding of the natural world, particularly for the biological sciences and physical systems. This sometimes involves paradigm shifts in engineering to meet the challenges of high-dimensional unstructured data, sparse labels, etc. To realize these goals, I currently work at the intersection of:

• Foundation models (FMs) — LLMs, multimodal models, etc.
• (Mechanistic) interpretability & explainability — understanding how models function and make decisions
• Efficient ML for long-range modeling — geometric priors, new forms of convolution, etc.

Questions include: How do FMs work internally? How should we probe them to understand their learning mechanisms? And how do these internals provide advanced generation capability? Can we exploit these internals to create new FM capabilities, either innate or equipped? Can we replicate FM internals cheaply for new model design?

And for science: How do FMs reason over long space- or time-horizons? Can FMs act as “data copilots” to help us understand key actors that distinguish (spatial, temporal, or spatio-temporal) systems from one another? Perhaps to localize binding sites in protein structures (below) or identify cell neighborhoods contributing to aggressive cancers (further below) with minimal supervision.

Thesis Work

My thesis work focuses on building data copilots to advance precision oncology: cancer subtyping, treatment planning, and prognostics. Classifying patients (e.g. cancer stage) with high fidelity is important for clinical decision support, but we also require “evidence” behind class decisions. This ML-generated evidence is known broadly as feature attribution as it can extract class-specific features/regions captured in a datum. My thesis aims to:

Evaluate feature attribution given partial contexts

In the realm of high-dimensional data, many ML models today are considered “partial-context” since they are limited by data throughput. Can such models still perform reliable feature attribution? Are some architectures or attribution methods inherently better at detecting class-specific regions? We design an evaluation framework for improved architecture + attribution selection [ECCV 22].

Build performant feature attribution methods for large models & data

We reason that feature attribution can be enabled by richer data representations, and thus seek to equip this capability to FMs. To both interface with FMs and operate over high-dimensional unstructured data (e.g. images, graphs, text), we developed interpretable convolutional architectures called "prospector heads" to mine for class-specific regions [ICML 23, arXiv 24].

Deploy feature attribution for knowledge discovery

With FMs and scalable feature attribution, we seek to improve cancer subtyping, treatment planning, and prognostics by discovering novel spatial biomarkers of cancer progression in patient biopsies (like the cell graph below). To do this, we are building data copilots for single-cell spatial biology, including proteomics datasets provided by GE Global Research.

Preprints & Publications

While my thesis research focuses on interpretable and explainable AI in the context of high-dimensional biomedical data, I've been privileged to work in several methods and application domains. A copy of my CV can be found at the bottom of this page. Stay tuned for future publications on:

1. Multimodal FMs that pass medical board exams
2. Data copilots for spatial biology: in silico discovery of biomarkers of cancer progression
3. Review paper: explainable ML for digital pathology & spatial biology
4. Characterizing human- and LLM-derived sequential decisions in competitive games like chess
5. VLM-driven zero-shot segmentation with long text captions
6. Fact-checking multimodal FMs to improve image tagging & retrieval
7. Satellite imagery VLMs fine-tuned to predict measures of maternal health & child mortality
8. Satellite imagery VLMs fine-tuned for few-shot detection of human labor trafficking encampments in the Amazon Rainforest

• Selected
• All

Vitæ

More details (projects, collaborators, talks, academic service, relevant coursework) can be found on my CV and LinkedIn page.

• Stanford University 2018 - present

  PhD Student
  Biomedical Informatics
  Data copilots for scientific discovery
• IBM Research Summer-Fall 2023

  Student Researcher
  Fact-checking Foundation Models
• Canary Center at Stanford Sep 2016 - Aug 2018

  Bioinformatics Research Assistant
  Mathematical modeling of tumors,
  NLP of biomedical literature,
  & anomaly detection in multi-omic time-series
• Strateos Summer 2016

  Bioengineering SDE Intern
  Software engineering for robotics:
  queueing, optimization, search
• UCB Center for Comp Bio Sep 2015 - June 2016

  Undergraduate Researcher
  Genomic sequence alignment algorithms
• Strand Life Sciences Summer 2015

  Bioinformatics Intern
  Genomic rare variant identification
• Lawrence Berkeley National Lab July 2013 – May 2015

  Undergraduate Researcher
  Computational Biophysics &
  structural biology of enzymes
• Berkeley BioLabs Summer 2014

  Bioengineering Intern
  Gene transfer experimentation
• University of California, Berkeley 2012 - 2016

  BA Student
  Applied Mathematics
  (emphasis in Mathematical Biology)
  Minor in Bioengineering

Process

Touching grass: On weekends, you can find me and my partner grabbing some brunch and sun in urban-suburban East Bay. Or we're unplugging at one of California's numerous regional, state, or national parks to spend time on the water and trails. Outside of hiking and camping, my hobbies include painting, gardening, and hosting regular themed dinner and cocktail parties. I also love city trekking with friends and popping into cafés and roasteries, museums, used bookstores (hunting for vintage maths sections), and outdoor pubs with live music sessions. Reach out if you'd like to grab a coffee!

Design: One of my favorite aspects of research is thinking about aesthetic and design when communicating technical ideas. This drive to understand ideas by visually communicating them (often to myself) sparked as a dyslexic Maths undergraduate. Despite my numerous interests in Maths, I struggled to parse and conceptualize blocks of textual abstraction in modern mathematical presentation, typical of standard teaching materials. I thus relied heavily on intuition and visual proofs as mental anchors. Thanks in part to training as a CIR Scholar at Stanford's Hasso Plattner Institute of Design, I cartoon-ify almost everything I work on and often spend Friday afternoons reflecting on, mocking up, and refining any discussed concepts.

• Selected
• All

Detecting salient objects
toy examples to show class differential features, presented in behind our ICML IMLH 2023 paper

Systems for Foundation Models
describing advances on chip design + distributed training

New architecture to learn salient objects
depicting graphical data structures behind our ICML IMLH 2023 paper

Nash Equilibria for joint optimizers
multiple choice grid depicted for each training step

Mpox mobile surveillance
graphical abstract behind our Nature Medicine (2023) paper

Increasing context lengths
comparison of lab's architectures

Detecting salient objects
toy examples to show class differential features, presented in behind our ICML IMLH 2023 paper

Systems for Foundation Models
describing advances on chip design + distributed training

New architecture to learn salient objects
depicting graphical data structures behind our ICML IMLH 2023 paper

Nash Equilibria for joint optimizers
multiple choice grid depicted for each training step

Mpox mobile surveillance
graphical abstract behind our Nature Medicine (2023) paper

Increasing context lengths
comparison of lab's architectures

Evaluating model explanations
graphical abstract behind our ECCV 2022 paper

Vision for instructive AI
talking points for where models could help humans

Global health monitoring
graphical abstract behind remote sensing for health surveillance

Hallmarks of cancer
background information on cancer progression

Visual grounding
desiderata to create more expressive Foundation Models

Physics-based model of cancer
graphical abstract for our Nature Sci Reports (2021) paper

Acknowledgement