Anastasiia Makarova

Anastasiia Makarova

PhD student

ETH Zurich

I work on sequential decision-making and representation learning for structured data, such as ​point clouds or graphs.

Research
The main question that motivates my research is: How can we actively learn new complex environments? My interests span multiple topics around sequential decision-making, bayesian optimization, representation learning, generative modeling. I am sparked by designing robust algorithms with quantified uncertainty and well-understood limitations (theoretically and empirically), that would be applicable in society-critical areas.

Previously, I was a doctoral student in Learning and Adaptive Systems Group at ETH Zurich supervised by Andreas Krause. My dissertation focuses on Bayesian optimization, proposing methods for risk-averse and computationally effective decision-making. Prior to that, I received master’s degrees in computer science and math from MIPT and Skoltech and a bachelor’s degree in math and physics from MIPT. In my master’s, I visited Columbia University in NYC and worked on deep learning for weakly-supervised semantic segmentation supervised by Victor Lempitsky and Hod Lipson.

I did research internships at Google DeepMind (RL and RLHF for LLMs) Yandex (deep learning for precipitation nowcasting) and Amazon Web Services (Bayesian optimization for AutoML).

Education
  • Dr. Sc., 2023

    ETH Zurich

  • MSc in Computer Science, 2017

    Skolkovo Institute of Science and Technology (Skoltech), Moscow

  • MSc in Applied Mathematics (with honors), 2017

    Moscow Institute of Physics and Technology (Phystech), Moscow

  • BSc in Applied Math and Physics (with honors), 2015

    Moscow Institute of Physics and Technology (Phystech), Moscow

News

  • October 2023: Safe risk-averse BO for controller tuning (paper) accepted to IEEE Robotics and Automation Letters.
  • September 2023: I successfully defended my doctoral thesis titled “Bayesian Optimization in the wild: risk-averse and computationally-effective decision-making”! Grateful to my advisor Andreas Krause, and to my brilliant collaborators!
  • August 2023: New preprint on arxiv: Adversarial Causal BO.
  • July 2023: New preprint on arxiv: Safe Risk-averse Bayesian Optimization for Controller Tuning.
  • May 2023: Attending ICLR in Rwanda and giving a talk about Model-based Causal BO, slides and talk
  • April 2023: Joining Google DeepMind, Brain team, as a research scientist intern, going to dive into RL from human feedback for LLM
  • January 2023: Model-based Causal BO accepted to ICLR featured spotlight (top 25% of accepted papers)
  • August 2022: Talk at Google TechTalk (Google BayesOpt Speaker Series), slides and video are available.
  • August 2022: Talk at AWS ML Science Tech Presentations series, slides are available.
  • Summer 2022: Our paper got the best paper award at AutoML Conf! I gave a contributed talk, recording, paper, intuitive blog post.

Publications

(2023). Adversarial Causal Bayesian Optimization.

Cite PDF

(2023). Safe Risk-averse Bayesian Optimization for Controller Tuning.

Cite PDF

(2022). Model-based Causal Bayesian Optimization. Spotlight at International Conference on Learning Representations (ICLR) 2023.

Cite Oral talk PDF Code Slides

(2022). Automatic Termination for Hyperparameter Optimization. Best Paper Award and Contributed Talk at International Conference on Automated Machine Learning (AutoML-Conf) 2022, ICLR Workshop on Neural Architecture Search 2021.

Cite Best paper award talk PDF Code Video Slides Blog post

(2021). Risk-averse Heteroscedastic Bayesian Optimization. Conference on Neural Information Processing Systems (NeurIPS).

Cite PDF Video Slides

(2021). Cherry-Picking Gradients: Learning Low-Rank Embeddings of Visual Data via Differentiable Cross-Approximation. International Conference on Computer Vision (ICCV).

Cite PDF Code Poster

(2020). Mixed-Variable Bayesian Optimization. International Joint Conference on Artificial Intelligence (IJCAI).

Cite PDF Short Video Long Video

(2020). Hierarchical Image Classification Using Entailment Cone Embeddings. CVPR Workshop on on Differential Geometry (DiffCVML), resulted into the publicly available mobile app BioDex.

Cite PDF Slides Code App

Experience

 
 
 
 
 
Research Scientist Intern
Google DeepMind
Aug 2023 – Apr 2023 Zurich
Research in RL and RLHF for LLMs.
 
 
 
 
 
ML Intern
Jun 2020 – Oct 2020 Berlin
Research Intern at AWS AI working on Automatic Model Tuning and SageMaker Autopilot.
 
 
 
 
 
PhD Student / Research Assistant
Oct 2017 – Present Zurich

Research areas: Probabilistic Machine Learning, Bayesian Optimization, Deep Learning, Tensors, Computer Vision

Advisor: Prof. Andreas Krause

 
 
 
 
 
Research Intern
Jun 2016 – Aug 2016 Moscow
Research Intern at Yandex.Weather working on precipitation nowcasting using deep convolutional and recurrent models
 
 
 
 
 
Visiting Researcher
Dec 2016 – Sep 2016 New York

Worked with Hod Lipson and collaboration with Victor Lempitsky.

  • Worked on deep learning based methods for weakly-supervised semantic segmentation
  • Developed an efficient architecture for image-based plant disease detection

Teaching

Supervised theses

I (co-)supervised MSc theses of several bright students, some resulting into research publications:

  • Alicja Chaszczewicz: Following Gradients to Calibrate Equilibrium Reaching Simulators,
    jointly with Max Paulus, ETH Zurich, December 2020 - May 2021.
  • Ankit Dhall: Learning Representations for Images With Hierarchical Labels (paper @DiffCVML'20),
    jointly with Octavian Ganea and Dario Pavllo, ETH Zurich, March - September 2019.
  • Erik Daxberger: Mixed-Variable Bayesian Optimization (paper IJCAI'20),
    jointly with Matteo Turchetta, ETH Zurich, October 2018 - April 2019.
  • Stefan Beyeler: Multi-fidelity Batch Bayesian Optimization for the Calibration of Transport System Simulations,
    jointly with Matteo Turchetta, ETH Zurich, October 2017 - April 2018.