Andrei Atanov

I am a PhD student at VILAB, EPFL, supervised by Amir Zamir. Previously I was a junior researcher at BayesGroup and Samsung-HSE Lab, supervised by Dmitry Vetrov. I received my bachelor's and master's degree in Computer Science from the HSE University.

I am broadly interested in understanding and building better deep learning model. My current research focuses on data-centric deep learning. Data is the fuel of deep learning, and I am interested in understanding how deep learning works through the lens of the training data and, more broadly, input signal. Leveraging this knowledge, I aim to develop algorithms that create better training data and, consequently, better models. Examples of research questions:

  • What is the optimal dataset for training a model that generalizes well?
  • What is the optimal perception signal for an agent to complete a given task?

๐Ÿ“ง Email  /  ๐Ÿ“‹ CV  /  ๐ŸŽ“ Google Scholar  /  ๐Ÿง‘โ€๐Ÿ’ป GitHub  /  ๐Ÿฆ Twitter

profile photo
Publications
Controlled Training Data Generation with Diffusion Models
Andrei Atanov*, Teresa Yeo*, Harold Benoit, Aleksandr Alekseev, Ruchira Ray, Pooya E. Akhoondi, Amir Zamir,

Arxiv, 2024 [Project Page] [arXiv: coming soon!] [GitHub: coming soon!]

We propose a method to generate synthetic training data specifically useful for a given supervised model and image domain. We introduce two feedback mechanisms to guide the generation: 1) model-based and 2) target distribution-based.

Unraveling the Key Components of OOD Generalization via Diversification
Andrei Atanov*, Harold Benoit*, Liangze Jiang*, OฤŸuzhan F. Kar, Mattia Rigotti, Amir Zamir,

ICLR, 2024 [paper]

We study the key components of diversification methods that were recently shown to achieve state-of-the-art OOD generalization. We show: 1) they are sensitive to the distribution of unlabeled data, 2) diversification alone is insufficient for OOD generalization and the right learning algorithm is needed and 3) the choice of the learning algorithm and the unlabeled data distribution are co-dependent.

Task Discovery: Finding the Tasks that Neural Networks Generalize on
Andrei Atanov, Andrei Filatov, Teresa Yeo, Ajay Sohmshetty, Amir Zamir,

NeurIPS, 2022 [Project Page] [arXiv] [GitHub]

What are the tasks a neural network can generalize on? How do they look, and what do they indicate? We introduce a task discovery framework to approach these questions empirically. These tasks reflect the inductive biases of NNs learned with SGD and can shed more light on how they generalize. Or do not! We show how one can use these tasks to reveal the failure modes of NNs via creating adversarial train-test data partitions.

Multimae: Multi-modal multi-task masked autoencoders
Roman Bachmann*, David Mizrahi*, Andrei Atanov, Amir Zamir,

ECCV, 2022 [Project Page] [arXiv] [GitHub]

A new pre-training method that captures and leverages cross-modal interactions.

3D Common Corruptions and Data Augmentation
OฤŸuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, Amir Zamir,

CVPR, 2022 [Oral] [Project Page] [arXiv] [GitHub]

3DCC is a set of more realistic 3D corruptions that can be used as a benchmark or data augmentations.

The Deep Weight Prior
Andrei Atanov*, Arsenii Ashukha*, Kirill Struminsky, Dmitry Vetrov, Max Welling,

ICLR, 2019 [GitHub]

We propose a flexible prior distribution over convolutional kernels of Bayesian neural networks.

Semi-Conditional Normalizing Flows for Semi-Supervised Learning
Andrei Atanov, Alexandra Volokhova, Arsenii Ashukha, Ivan Sosnovik, Dmitry Vetrov,

INNF Workshop at ICML, 2019 [GitHub]

We apply conditional normalizing flows to semi-supervised learning. We utilize multiscale architecture for computational efficiency.

Uncertainty Estimation via Stochastic Batch Normalization
Andrei Atanov, Arsenii Ashukha, Dmitry Molchanov, Kirill Neklyudov, Dmitry Vetrov,
ICLR Workshop Track, 2018

We propose a probabilistic view on Batch Normalization and an efficient test-time averaging technique for uncertainty estimation in batch-normalized DNNs.


Credits to Jon Barron for the template.