Deep Learning · Robustness · Efficiency . Histopathology

Dhananjay Tomar

I build robust and efficient deep-learning systems for real-world data.

I am a researcher at the Institute for Cancer Genetics and Informatics (ICGI), Oslo University Hospital (OUS), and a PhD candidate at the University of Oslo (UiO). My work spans computer vision and histopathology, with a focus on robustness, efficiency, and the training dynamics of neural networks.

Deep Learning researcher Oslo, Norway

Experience

Research, teaching, and production machine learning.

Oct 2025 — Present

Scientific Developer (research role) @ ICGI

Oslo University Hospital

Researching cancer-survival prediction, including a label-denoising method and methods for making models more robust to scanner differences.

Oct 2021 — Present

PhD Candidate @ University of Oslo

Researching robust and efficient computational pathology, including NeurIPS 2024 work on nuclear-shape-guided learning; two additional manuscripts are under review.

I co-supervised two master's students and, for nearly four years, taught the course Deep Learning for Image Analysis, where I lectured, designed mandatory assignments, contributed to exams and marking, supported students, and helped run the course.

Jan 2018 — Sep 2021

Machine Learning Developer @ SAP

Designed major components of the Document Information Processing service that processed over a million financial documents per month. My work spanned the full spectrum of applied ML—from delivering a successful accruals automation POC to developing production algorithms and patenting a novel synthetic data generation method.

Selected research

Published work and earlier projects; manuscripts under review stay anonymous here.

2024

Published NeurIPS 2024

Are nuclear masks all you need for improved out-of-domain generalisation?

Uses nuclear masks during training to steer a cancer classifier toward nuclear shape and organisation, while requiring only the original image at inference; improves transfer across hospitals and robustness to corruptions and adversarial attacks.

2017

Thesis MSc thesis · IDSIA / USI

Curriculum learning and neural-network training dynamics

Tracked examples through training using time-to-learn, loss changes, and computation selected by Highway Network gates. The experiments exposed multiple notions of difficulty and showed that a batch's effect changes with the model's training state.

2017

Poster PASC17

Pruning Highway Networks for reducing memory usage

Turned transform gates into pruning signals, reducing a CIFAR-10 Highway Network from 34 MB to 5.2 MB—about 85% smaller—with a 1.72 percentage-point accuracy drop.

2017

Published IEEE conference paper

Feature Selection Using Autoencoders

Introduced a feature-selection method that filters uninformative inputs by tracing high-variance connections backwards through an autoencoder; evaluated on five public datasets, including 59% pixel removal on MNIST with roughly a 1% accuracy loss.

Patent & service

Selected contributions beyond the main publication list.

Reviewing

NeurIPS 2025 Top Reviewer Award

Recognized as a Top Reviewer (Top 10%) for delivering high-quality, timely paper evaluations.

Official Top Reviewer List

Patent · US 20230334309 A1

Synthetic training data for document extraction

Conceived the core idea for a synthetic training-data generation method that applies macro and micro augmentations to electronic-document templates, generating new documents for training ML models.

View patent

Education

Formal training across India, Switzerland, and Norway.

2021 — Present

PhD in Informatics

University of Oslo

Robust and efficient deep learning for computational pathology.

2015 — 2017

MSc in Informatics

Università della Svizzera italiana

Master's thesis at IDSIA on curriculum learning and training dynamics.

2011 — 2015

BTech in Computer Science and Engineering

Jaypee Institute of Information Technology

Major project on neural-network-based feature selection.

Latest writing

Research stories, paper walkthroughs, and notes.

Research, collaboration, or a good paper recommendation

Let's talk.

Email Dhananjay