We study the problems of generative models: time frame-by-frame connectivity of the generated video, convergence of models during training in denoising tasks, interpretability and controllability of generation processes within codecs.
We create diagnostic methods for complex systems for timely and preventive impact.
We investigate complex architectures (img2img, Transformer) for detection, segmentation, resolution enhancement, 3D reconstructions, etc.
Zero- and Few-shot approaches
We solve problems with a significant lack of data: there are several examples in the sample with a large number of classes. We apply small-choice approaches to computer vision tasks (video stream detection and segmentation) and natural language processing (text classification and tagging).
We study the applicability of AI methods to encoding information of various nature: from the video stream to the human genome.
We create deep learning models that can be interpreted and managed in the process of use.
Clients of research teams form the goals of R&D projects. We have highlighted the most common research requests:
Preliminary and detailed study of the subject area with the search for scientific articles, available data, and open source in order to obtain a base of materials for a quick dive into a new area and form requirements for future R&D results
Technological Pipeline Development
Development of a complete technological pipeline for solving a problem based on client data, based on existing algorithms and open source in order to obtain a strong basic solution for subsequent modifications and growth of target metrics
Negative Effect Removal
Elimination of a key negative effect in the operation of the existing model: from excessive sensitivity to fluctuations in the data to the quality of work in corner cases. The goal is to grow certain metrics of the customer model within the given constraints
Industrial Pipeline Development
Development of a Proof-of-the-Concept pipeline for solving an industrial business problem based on artificial intelligence algorithms in the absence of known approaches to its solution. The goal is the fundamental possibility of solving such a problem
The methodology for the implementation of R&D projects developed on the basis of our experience allows us to quickly carry out a full cycle of research and reduce risks