I am mainly interested in solving new problems with Natural Language Processing models and making current advances in deep learning more efficient and broadly accessible. In particular, my research on decentralized deep learning has served as a basis of Hivemind, a PyTorch library for training neural networks over heterogeneous, unreliable and poorly connected hardware. Before that, I also worked on topics like uncertainty estimation in machine translation, graph-based word representations, neural networks with adaptive computation, and gradient optimization of decoding hyperparameters in language generation.