MIL Team Research

We create new models of Artificial Intelligence

We write scientific articles

We develop compression methods for neural network models: distillation, pruning, quantization, search for neural network architectures

Generative models

We study the properties of generative models for creating text, images and video: frame-by-frame coherence, model convergence, audio denoising, interpretability and controllability of generation.

AI-codecs

We build models for encoding and decoding audio, video and other information using neural network models. We study the influence of the semantics of the encoded object on the quality of codecs.

Zero- and Few-shot approaches

We create methods applicable when there is a significant lack of data: several examples with a large number of classes - for the fields of computer vision and natural language processing.

Interpretable models

We create deep learning models that are interpretable and manageable. We explore the properties of the constructed models, limitations and areas of application.

Audio Analysis

We use generative and discriminative deep learning approaches to denoise audio tracks, change their style, and detect emotions.

Wearable devices

We apply algorithms to data from inertial and biomedical sensors to solve problems in HAR, pressure determination, navigation, and so on.

NLP Analytics

We create a full set of text analysis methods: hard and soft clustering, summarization, classification and tagging, search tasks. We use large language models and classical methods.

Computer vision

We explore complex architectures (img2img, Transformer) for detection, segmentation, resolution enhancement, 3D reconstruction, and so on.

Predictive Analytics

Создаем методы диагностики сложных систем для своевременного и превентивного воздействия.

Completed research projects

Published scientific articles

What projects do we help corporations with:

Preliminary Research
Exploratory analysis of a new topic: searching for scientific articles, available data, open source code, forming hypotheses.
The result is a base of materials for quick immersion in the area and requirements for future stages.
Technical pipeline development
We assemble a pipeline from a variety of existing algorithms for solving a problem based on client data.
The result is a library with a basic solution and a list of bottlenecks for subsequent modifications.
Remove negative effect
We eliminate the key negative effect in the model’s operation: from instability to low quality in target conditions.
The result is a new version of the model that meets the specified criteria.
Development of an industrial pipeline
We come up with and create a prototype for solving a client’s business problem based on modern AI models. The result is a prototype demonstrating the fundamental possibility of solving the problem.
Increase in quality
We collect samples and further train models to achieve a given quality threshold using client data.
The result is a new version of the model that meets the target metrics.
Model compression
We compress models programmatically to meet the requirements for computational complexity at a given quality.
The result is a new version of the model that meets the target requirements.

Flexibility

Flexible approach to project scope management and transparency of the development process

Methodology

Using AI project methodology to reduce risks and systematically achieve breakthrough results

Qualification

Deep understanding of both core business needs and the use of SotA AI methods

Iteration

Short iterations with the transfer of a useful and exploitable result

Scientific consultants Research Group

Scientific consultants of the Research Group

Konstantin Vorontsov
Prof, MSU
Specialisation: NLP
Andrey Leonidov
Prof, CERN
Specialisation: ML
Vadim Strijov
Prof, Grenoble
Specialisation: Sensors
Andrey Raygorodsky
Prof, Yandex
Specialisation: ML, Graphs
Mikhail Burtsev
Prof, AIRI
Specialisation: NLP
Radoslav Neychev
Consult
Specialisation: Sensors
Ilya Zharikov
Consult
Specialisation: DL & Sensors
Oleg Bakhteev
Consult
Specialisation: DL