NN Architecture comparison
Overview of the field: methods for fast and accurate comparison of the quality of neural network architectures that do not require a full training procedure, which usually requires a large amount of computing resources
Project with the company
Huawei
DESCRIPTION

The project team had to conduct a full review of the scientific literature on methods for effective training of neural network models, evaluating their quality, and ranking models according to their quality within the framework of the NAS task, which allow reducing the number of computing resources used several times.

RELEVANCE

Methods of ranking models according to their quality significantly speed up the selection of architecture for automatic search, reducing the full search procedure to several GPU hours. Methods of effective training and quality assessment allow you to quickly get the final quality of the model, which leads to faster development of the final model and reduces the amount of computing resources needed to solve the problem.
The decision of the team
The laboratory team analyzed more than 70 works, based on 13 of which they formulated a project vision and formed a work plan for implementing the required models.
Results
- Compiled an overview of the area
- Selected areas of model development
Team
- Team Lead: Ilya Zharikov
- Project Manager: Victoria Yanaeva
- Research team: Philip Nikitin, Ivan Krivorotov