Samyak Jain


Interested in AI Safety (Looking for PhD positions)

Hi! I am Samyak. I am currently working at Five AI at Oxford, where I am advised by Puneet Dokania. Previously, I have worked with David Krueger at Cambridge University and Venkatesh Babu at Indian Institute of Science, Bangalore. Recently, I completed my bachelors and masters in Computer Science at the Indian Institute of Technology (BHU) Varansi, where my thesis focussed on understanding and improving adversarial robustness of neural networks. For the last few years I have been broadly working on Machine Learning with a particular focus on Adversarial Robustness, Generalization and Mechanistic Interpretability. Moving forward, I am excited about contributing in the field of AI Safety.

Recently, I completed my interenship at Cambridge University, where I worked on mechanistic interpretability. Before that I worked at the Vision and AI Lab where my projects were related to adversarial robustness and understanding loss landscape of neural networks, which have been published at the following venues.

Besides this I actively participate in machine learning and computer vision conferences and have served as a reviewer at multiple conferences and won outstanding reviewer awards at ICLR 2022 and CVPR 2022.

Here is the link to my CV.

Broadly I am interested in research topics related to AI safety, in particular:
Alignment: Unlearning, RLHF fine-tuning
Mechanistic Interpretability: Science of Deep Learning
Robustness: Adversarial Robustness, OOD Robustness
Generalization and Foundations of ML: Mode Connectivity, Loss landscape sharpness, Learning dynamics
Role of Data in Learning: Curating data to build safe models


Jan 15, 2024 Our work mechanistically analyzing the effects of fine-tuning on procedurally defined tasks was accepted at ICLR 2024.
Jun 5, 2023 Glad to be recognized as Outstanding Reviewer (top 250 reviewers) at CVPR 2023.
Jun 1, 2023 Excited to start a research internship at CBL lab at Cambridge University under Prof. David Krueger
Mar 1, 2023 Our work titled “DART: Diversify-Aggregate-Repeat Training Improves Generalization of Neural Networks” got accepted at CVPR 2023. The work is available online. :sparkles: :smile:
Sep 16, 2022 Our work on “Efficient and Effective Augmentation Strategy for Adversarial Training” got accepted at NeurIPS 2022. The work is available online. :sparkles: :smile:
Aug 2, 2022 Our work on “Scaling Adversarial Training to Large Perturbation Bounds” got accepted at ECCV 2022. The work is available online. :sparkles: :smile:
Jul 28, 2022 Glad to be recognized as Outstanding Reviewer (top 200 reviewers) at CVPR 2022.
Apr 20, 2022 Glad to be recognized as Highlighted Reviewer at ICLR 2022.
May 5, 2021 Was Awarded the prestigious DAAD WISE Scholarship for a research intern at a German University.
Apr 21, 2021 Started working as a research intern at Theoretical foundations in AI at the Technical University of Munich.