Skip to content

About

Aaditya (Adi) Prakash

  • Principal ML Scientist
  • AI x Healthcare
  • PhD Machine Learning

Google Scholar LinkedIn GitHub

Biography

I am a Machine Learning (Deep Learning) scientist with interests in applications of computer vision in healthcare and life sciences. I am a Principal Machine Learning Scientist at Genentech, which I joined in January 2025 when Genentech acquired Spring Science (formerly Spring Discovery); I build single-cell embeddings and foundation models for high-content imaging and optical pooled screens. At Spring I developed foundation models for drug imaging data and ML features used by biologists and by several big pharma partners. Previously, I was at PathAI as a Manager and Research Lead building state-of-the-art, interpretable multiple-instance models for digital pathology. I have a Ph.D. in Computer Science from Brandeis University with a specialization in applications of deep learning in vision and language, an M.A. in Computer Science from Brandeis, and a B.S. in Biomedical Engineering. In the past, I have worked at Microsoft Research, Qualcomm Research, and Philips Research.

I am from Kathmandu, Nepal. Other than research, I love rock climbing, running, playing chess and listening to classical Indian songs.

Experience

Principal Machine Learning Scientist

Genentech (Roche Group)

January 2025 – Present · Boston, MA

  • Single-cell embeddings for high-content imaging and optical pooled screens.
  • Developing various AI tools for drug discovery research.
  • Joined with the acquisition of Spring Science (formerly Spring Discovery), whose AI R&D platform was acquired by Genentech, a leading biotech in the Bay Area.

Staff Machine Learning Scientist

Spring Discovery → Spring Science

May 2022 – December 2024 · Boston, MA

  • Developed foundation models for drug imaging data with a 5× improvement over off-the-shelf models.
  • AI features (embeddings) for drug similarity and toxicity screens used by several big pharma companies.
  • Identify prototype single cells for drug phenotypes and MoAs.
  • Developed and published methods for removing confounders in medical imaging (SLAS 2023).
  • Designed and implemented similarity features for the Spring Science app used by biologists and data scientists.
  • Developed an efficient backend for embedding generation over billions of cells to enable biologists to use SOTA AI.

Manager and Research Lead, Machine Learning

PathAI

January 2019 – May 2022 · Boston, MA

  • Investigated new avenues (patient triaging, second-read) for applications of AI in digital pathology.
  • Developed state-of-the-art techniques in multiple instance learning for whole slide images - published Additive MIL (NeurIPS 2022).
  • Explored solutions for model training under limited labels (self-supervision) and label noise.
  • Managed a team of machine learning engineers to enable new ML capabilities at PathAI; developed and productionized efficient models for cell detection and classification.
  • Predicted drug response in breast cancer (ESMO 2020), fibrosis progression from MRI (EASL 2021), and molecular phenotypes from human-interpretable WSI features (Nature Comm. 2021).

Research Scientist, AI Labs (intern in 2016)

Philips Research North America

August 2016 – January 2019 · Cambridge, MA

  • Demonstrated that neural networks can detect adverse drug reactions (WWW 2017).
  • Developed methods for clinical paraphrase generation and clinical text simplification using unsupervised models (COLING 2016; multiple patents).
  • Explored applications of LSTM in sequence-to-sequence learning (COLING 2016).
  • Developed an efficient representation of memory state for memory networks (AAAI 2017).

Research Intern, AI + Research

Microsoft Research

Summer 2018 · Redmond, WA

  • Model compression in convolutional neural networks - RePr (CVPR 2019, Oral).
  • Improved training of compact models (MobileNet, SqueezeNet, ShuffleNet).
  • Developed a new technique for efficient training of neural networks.

Research Intern

Qualcomm Research

Summer 2017 · San Diego, CA

  • Explored model parallelism for convolutional neural networks.
  • Architecture learning for reduced model complexity.
  • Developed custom model compression for the Neural Processing Engine.
  • Received the Roberto Padovani Scholarship (Qualcomm).

Big Data Analyst & Teaching Assistant

Brandeis University

2013 – 2019 · Waltham, MA

  • Researched new techniques in data analysis on Hadoop and Spark frameworks.
  • Teaching assistant for Mobile App Development, Scientific Computing, Artificial Intelligence, Big Data Analysis, Theory of Computation, Data Structures, Algorithms and Data Compression.

Publications

first author marks first-authored work. Full list with citations: Google Scholar. Posters and slides: research-papers-slides-posters.

AI powered quantification of nuclear morphology in cancers enables prediction of genome instability and prognosis

npj Precision Oncology 2024

J. Abel, S. Jain, D. Rajan, H. Padigela, K. Leidal, A. Prakash, J. Conway, et al. - AI features from pathology whole-slide-images to predict various genomic conditions.

[PDF](https://pmc.ncbi.nlm.nih.gov/articles/PMC11187064/) · [PubMed](https://pubmed.ncbi.nlm.nih.gov/38898127/)

ContriMix: Disentanglement of content & attribute for domain generalization in microscopy images

2024

T. H. Nguyen, D. Juyal, J. Li, A. Prakash, S. Nofallah, C. Shah, S. C. Gullapally, et al. - Domain generalization by disentangling biology and attributes.

[arXiv](https://arxiv.org/abs/2306.04527)

Synthetic DOmain-Targeted Augmentation (S-DOTA) improves model generalization in digital pathology

2023

S. C. Gullapally, Y. Zhang, N. K. Mittal, D. Kartik, S. Srinivasan, K. Rose, … A. Prakash, et al. - Identifying and improving generalization issues across imaging scanners.

[arXiv](https://arxiv.org/abs/2305.02401)

Cell-type-specific nuclear morphology predicts genomic instability and prognosis in multiple cancer types

bioRxiv 2023

J. Abel, S. Jain, D. Rajan, H. Padigela, K. Leidal, A. Prakash, J. Conway, et al. - Nuclei features that help identify cancer-causing genotypes (BRCA, LUAD and PRAD).

[PDF](https://www.biorxiv.org/content/10.1101/2023.05.15.539600)

Additive MIL: Intrinsically Interpretable Multiple Instance Learning for Pathology

NeurIPS 2022

S. A. Javed, D. Juyal, H. Padigela, A. Taylor-Weiner, L. Yu, A. Prakash - MIL method that generates visual attention which is grounded and correlates with segmentation.

[PDF](https://arxiv.org/pdf/2206.01794) · [arXiv](https://arxiv.org/abs/2206.01794)

AI-powered segmentation and analysis of nuclei morphology predicts genomic and clinical markers in multiple cancer types

AACR 2022

J. Abel, S. Jain, D. Rajan, K. Leidal, H. Padigela, A. Prakash, J. Conway, et al. - Nuclei segmentation and morphological features of nuclei are predictive of certain cancers.

A deep learning approach to analysis of MRCP images predicts clinical events and progression to cirrhosis in patients with primary sclerosing cholangitis

EASL 2021 first author

A. Prakash, et al. - An application of a two-stage ConvNet to predict cirrhosis using MRI images.

Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular phenotypes

Nature Comm. 2021

- Application of ConvNets to find features that are interpretable and predict molecular signal on WSI.

Artificial intelligence analysis of advanced breast cancer patients from a phase I trial of trastuzumab deruxtecan (T-DxD): HER2 and histopathology features as predictors of clinical benefit

ESMO 2020

- An application of ConvNet and extracted features to predict drug response.

RePr: Improved Training of Convolutional Filters (Oral)

CVPR 2019 first author

A. Prakash, et al. - Improved performance of vanilla CNNs without using residual or dense connections.

[PDF](https://arxiv.org/pdf/1811.07275) · [Code](https://github.com/iamaaditya/repr-ortho)

Compact Representations of Dynamic Video Background Using Motion Sprites

IEEE DCC 2019

- Technique to store video background motions as a time-invariant representation of the optical flow.

Deflecting Adversarial Attacks with Pixel Deflection (Spotlight)

CVPR 2018 first author

A. Prakash, et al. - Image transformation based defense to adversarial attacks; recovers 98% of fooled images.

[PDF](https://arxiv.org/pdf/1801.08926) · [Code](https://github.com/iamaaditya/pixel-deflection) · [Video](https://www.youtube.com/watch?v=VgjOXJ9QKWo)

Robust Discriminative Localization Maps

CV-COPS 2018 first author

A. Prakash, et al. - Securing class activation maps against attacks by using the geometric mean over all classes.

[Code](https://github.com/iamaaditya/robust-activation-maps)

Protecting JPEG Images Against Adversarial Attacks (Oral)

IEEE DCC 2018 first author

A. Prakash, et al. - Improves the ability of JPEG to defend against attacks; recovery improved from 27% to 82%.

[PDF](https://arxiv.org/pdf/1803.00940) · [Code](https://github.com/iamaaditya/protecting-jpeg)

DR-BiLSTM: Dependent Reading Bidirectional LSTM for NLI

NAACL 2018

- Dependent reading using hierarchical soft attention; achieves SOTA on Stanford NLI.

Visual Lecture Summary using Intensity Correlation Coefficient

IMVIP 2017

- Technique to remove the instructor and generate slides from white/chalk board videos.

Condensed Memory Networks for Clinical Diagnostic Inferencing

AAAI 2017 first author

A. Prakash, S. S. Al Hasan, O. Farri, K. M. Y. Lee, V. Datla, A. Qadir, J. Liu - Classifying the diagnosis of a given medical note; SOTA results.

[PDF](https://arxiv.org/pdf/1612.01848v1.pdf) · [Slides](https://docs.google.com/presentation/d/1NSGEBYJmYEa5zsPJBocYcyRVZLKG0y274Wib_BlRX-U/edit?usp=sharing)

Semantic Perceptual Image Compression using Deep CNNs (Oral)

IEEE DCC 2017 first author

A. Prakash, et al. - Using custom designed CNNs to add differential quantization to achieve semantic JPEG.

[PDF](https://arxiv.org/pdf/1612.08712v1.pdf) · [Code](https://github.com/iamaaditya/image-compression-cnn)

Adverse Drug Event Detection in Tweets with Semi-Supervised CNNs

WWW 2017

- Use of unlabeled data to improve performance of detecting ADE in tweets; SOTA results on PSB 2016.

Neural Paraphrase Generation with Stacked Residual LSTM Networks

COLING 2016 first author

A. Prakash, et al. - First deep learning based paraphrasing model; use of skip connections on LSTM.

[PDF](https://arxiv.org/pdf/1610.03098v3.pdf) · [Code](https://github.com/iamaaditya/neural-paraphrase-generation)

Highway Networks for Visual Question Answering (honorable award)

CVPR (VQA) 2016 first author

A. Prakash, J. Storer - VQA model with implicit attention; top-4 in VQA Challenge 1.0.

[PDF](https://docs.google.com/viewer?url=https://github.com/iamaaditya/research-papers-slides-posters/raw/master/cvpr_2016/cvpr_paper.pdf) · [Slides](https://docs.google.com/presentation/d/1nrc-kOmDn1Msp41q8gwiiDFrXa0uKoBatlk1plmo8GM/edit?usp=sharing)

Patents

  • US 12,204,854 - System for multi-perspective discourse within a dialog (2025)
  • US 12,183,105 - Method to improve instance selection in bootstrapping framework for concept extraction from text documents (2024)
  • US 11,868,720 - System for multi-perspective discourse within a dialog (2024)
  • US App. 18/660,563 - Additive multiple instance learning (2024)
  • US 11,621,075 - Systems, methods, and apparatus for diagnostic inferencing with a multimodal deep memory network (2023)
  • US 11,544,529 - Semi-supervised classification with stacked autoencoder (2023)
  • US App. 18/130,517 - Condensed memory networks (2023)

Research highlights

  • Adversarial robustness

    Defenses against imperceptible adversarial perturbations without retraining: pixel deflection + wavelet denoising (CVPR 2018), and an adaptive JPEG encoder (DCC 2018). Robust discriminative localization maps (CV-COPS 2018).

  • Interpretable ML for pathology

    Additive MIL (NeurIPS 2022): exact spatial credit assignment for whole-slide models; AI quantification of nuclear morphology predicting genome instability and prognosis (npj Precision Oncology 2024).

  • Foundation models for single-cell biology

    Foundation models and single-cell embeddings for drug imaging data at Spring/Genentech with 5× gains over off-the-shelf models; PBMC inflammasome benchmark datasets and confounder removal (SLAS 2023).

  • Memory & language networks

    Condensed memory networks for clinical diagnostic inferencing (AAAI 2017); stacked residual LSTMs for paraphrase generation (COLING 2016); dependent reading BiLSTM for NLI (NAACL 2018).

Computing skills

  • Languages - Python, C, C++, CUDA
  • Deep Learning - PyTorch, TensorFlow, Keras, HuggingFace, DINOv2, SAMv2
  • Research tools - SciPy, NumPy, OpenCV, Git, Bash, LaTeX

Education & honors

  • Ph.D., Computer Science

    Brandeis University, 2015 – 2019. Advisor: Prof. James Storer. Thesis: Robust and Efficient Techniques in Deep Learning with applications in Computer Vision and Language Understanding.

  • M.A., Computer Science

    Brandeis University, 2013 – 2015. GPA 4.0/4.0. Algorithms, distributed systems, statistical approaches to NLP, computational semantics, computational neuroscience, computational biology.

  • B.S., Biomedical Engineering

    Bharath University, India, 2004 – 2008. GPA 9.36/10, rank 1/71. Gold medal for securing the highest rank.

  • RL Summer School

    Vector Institute (CIFAR/MILA), 2018. Lectures by Yann LeCun, Yoshua Bengio and Richard Sutton.

  • Best Reviewer Award, NeurIPS (2018)
  • Roberto Padovani (Qualcomm) Scholarship Award (2017)
  • Outstanding Teaching Fellow, Brandeis University (2017)
  • Honorable Spotlight Award, Visual Question Answering Challenge, CVPR (2016) [Video]
  • Best Paper Award, International Conference on Perspective of Computer Confluence, Pune (2012)
  • Gold Medal (highest rank), Bharath University, Chennai (2008)

Teaching & talks

  • CS 175 - Deep Learning lecture series (Brandeis): neural networks, CNNs, RNNs, object localization/detection, memory networks - slides & PDFs.
  • Seminar talks and class slides on deep learning, computer vision and GPGPU computing - presentations index.
  • Invited talk on Deep Learning at Connecticut College (2018) [Slides].
  • Posters and slides for all publications live in the research-papers-slides-posters repository.

Projects & code

Open source code and demos - full list on GitHub: