Mgr. Oleksandr Shekhovtsov, Ph.D.

The topic is at the intersection of machine learning and computer vision. You will research methods and models for handling different kinds of discreteness in modern statistical machine learning: discrete variables, structures, representations, training with non-differentiable losses; and apply them in computer vision. The main application of interest is learning Quantized Neural Networks (QNN), where weights and activations are represented with a few bits only. This offers huge savings in terms of computation cost and energy and allows larger models to run in simpler devices [1,2,5,6]. The challenge is to learn such quantized models to achieve high efficiency and accuracy. The research focuses on stochastic relaxation methods [3,4]. To quantize modern architectures, we need to develop suitable discretizations of weight kernels, residual connections, representations such as queries and keys in the attention model underlying transformers, etc. Further applications could be in conditional computing models and in architecture search. It can also be desirable to learn discrete representations on the output of a neural network. For example, for the image retrieval application we want to learn compact binary descriptors, which are efficient to store and fast to compare, such that similar descriptors correspond to semantically similar objects (contrastive learning). These representations can be latent, as in variational autoencoders and even structured (latent parse tree structures [7]). Training with non-differentiable losses is important e.g. for minimizing risk in statistical decision making [8].