Senya Ashukha

Arsenii Ashukha

I am a Senior Research Scientist at Isomorphic Labs, Alphabet company led by Demis Hassabis, where I work on AI-first drug design.

Before joining Isomorphic, I worked at Samsung AI and did my PhD with Dmitry Vetrov, in collaboration with Max Welling. The results of my PhD contributed to sparsification, uncertainty, and generative models.

Email / CV / Google Scholar / GitHub / Twitter / 📍London (UK)

Research

	Resolution-robust Large Mask Inpainting with Fourier Convolutions Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, Victor Lempitsky WACV, 2022 project page / poster video (5 mins) / arXiv / code / bibtex / Yannic Kilcher / Casual GAN LaMa generalizes surprisingly well to much higher resolutions (~2k❗️) than it saw during training (256x256), and achieves the excellent performance even in challenging scenarios, e.g. completion of periodic structures.
	Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning Arsenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, Dmitry Vetrov ICLR*, 2020 blog post / poster video (5mins) / code / arXiv / bibtex The work shows that i) a simple ensemble of independently trained networks performs significantly better than recent techniques ii) a simple test-time augmentation applied to a conventional network outperforms low-parameters ensembles (e.g. Dropout) and also improves all ensembles for free iii) comparison of uncertainty estimation ability of algorithms is often done incorectly in literature.
	Greedy Policy Search: A Simple Baseline for Learnable Test-Time Augmentation Arsenii Ashukha, Dmitry Molchanov, Alexander Lyzhov, Yuliya Molchanova, Dmitry Vetrov UAI, 2020 code / arXiv / slides / bibtex We introduce greedy policy search (GPS), a simple but high-performing method for learning a policy of test-time augmentation.
	The Deep Weight Prior Arsenii Ashukha, Andrei Atanov, Kirill Struminsky, Dmitry Vetrov, Max Welling ICLR, 2019 code / arXiv / bibtex The deep weight prior is the generative model for kernels of convolutional neural networks, that acts as a prior distribution while training on new datasets.
	Variance Networks: When Expectation Does Not Meet Your Expectations Arsenii Ashukha, Kirill Neklyudov, Dmitry Molchanov, Dmitry Vetrov ICLR*, 2019 code / arXiv / bibtex It is possible to learn a zero-centered Gaussian distribution over the weights of a neural network by learning only variances, and it works surprisingly well.
	Semi-Conditional Normalizing Flows for Semi-Supervised Learning Andrei Atanov, Alexandra Volokhova, Arsenii Ashukha, Ivan Sosnovik, Dmitry Vetrov INNF Workshop at ICML, 2019 code / arXiv / bibtex We employ semi-conditional normalizing flow architecture that allows efficiently trains normalizing flows when only few labeled data points are available.
	Unsupervised Domain Adaptation with SharedLatent Dynamics for Reinforcement Learning Evgenii Nikishin, Arsenii Ashukha, Dmitry Vetrov BLD Workshop at NeurIPS, 2019 code / poster Domain adaptation via learning shared dynamics in a latent space with adversarial matching of latent states.
	Uncertainty Estimation via Stochastic Batch Normalization Andrei Atanov, Arsenii Ashukha, Dmitry Molchanov, Kirill Neklyudov, Dmitry Vetrov Joint Workshop Track at ICLR, 2018 code / arXiv Inference-time stochastic batch normalization improves the performance of uncertainty estimation of ensembles.
	Structured Bayesian Pruning via Log-Normal Multiplicative Noise Kirill Neklyudov, Dmitry Molchanov, Arsenii Ashukha, Dmitry Vetrov NeurIPS, 2017 code / arXiv / bibtex / poster The model allows to sparsify a DNN with an arbitrary pattern of spasticity e.g., neurons or convolutional filters.
	Variational Dropout Sparsifies Deep Neural Networks Arsenii Ashukha, Dmitry Molchanov, Dmitry Vetrov ICML, 2017 retrospective⏳ / talk (15 mins) / arXiv / bibtex / code (theano, tf by GoogleAI, colab pytorch) Retrospective: SparsesVD works in practice and it was used for network sparsification in leading IT companies. However, future studies showed that careful usage of pruning-based methods can produce better results. Training of deep models with noise is known to be hard and unstable. That is less the case with SparseVD. All variances are initialized with small values and did not change much during training. Using small variances does not hurt the performance, thus SparseVD might be considered as a fancy regulariser with (almost) no noise. The sparse solution is just a local optimum, as better values of ELBO can be achieved with a less flexible variational posterior q(w_ij)=N(w_ij \| 0, σ_ij). Variational dropout secretly trains highly sparsified deep neural networks, while a pattern of sparsity is learned jointly with weights during training.