MISiS: Technology for selecting a neural network for face recognition

The main articles are:

• Neural networks (neural networks)
• Machine Learning

2023: Introduction of facial recognition application matching technology for electronics

Scientists at MISIS University, HSE and AIRI have proposed neural network technology, which, when installed on a smartphone, tablet or smart home device, analyzes their technical characteristics and selects the most optimal version of the face recognition application, which will work best on each specific device. This problem is especially true for manufacturing companies that have large batches of equipment with various microchips and characteristics. They need an accurate answer, which neural network for face recognition is better to put on a separate model of the device. MISIS announced this on October 10, 2023.

Quickly choosing the best neural network architecture for a specific mobile device is a difficult task. The computing power of cheap and expensive smartphones varies significantly, so it is impossible to find a single universal neural network with high accuracy and acceptable performance for all devices.

For October 2023, there are different architectures, but they cannot be adapted for a specific device. The advantages of this technique are that it is easy to integrate into any technique and does not need to be trained from scratch. The device will send information about the operating time of each layer of the neural network to the server, and it will send an optimal model for it in response.

When installed on a smartphone, a special application analyzes the technical capabilities of a specific device, and then the most accurate subnet is selected from the trained neural network SuperNet using the proposed algorithm, which will analyze the face image in a given time on this device. The presence of an already trained neural network, several subnets and a demonstration application for Android facilitates the practical implementation of the proposed framework, "says study co-author Ilya Makarov, director of the artificial intelligence center NUST MISIS, head of the AI in Industry group of the Institute of Artificial Intelligence AIRI.

One of the most difficult tasks pattern recognition is the tasks of checking and identifying persons. In typical scenarios, the training set contains a small number of photos for each person of interest. For October 2023, these problems are solved by extracting features, or descriptors, using a deep neural network previously trained on large external sets. Unfortunately, data even modern facial descriptors are characterized by racial bias, low accuracy in low light and often require repeated. identifications It is almost impossible to train a universal face descriptor that could be used for real-time facial recognition with high accuracy across all devices. One of the potential solutions is to use a neural architecture search engine to correctly select a neural network for a specific device and automatic (machine learning AutoML) methods.

We did not make a product, but an open technology that everyone can use. Imagine that you have lots of a million tablets on which you need to install a module for identifying faces. They have specific microchips and certain technical capacities. You can take one device out of the box, install our application, which will determine the most suitable model, which will be suitable for this particular tablet model and will recognize faces in 5.10 or 20 milliseconds. By the way, we solve the problem of not only identification, but also verification of persons, - explained the head of the study Andrei Savchenko, deputy director of the artificial intelligence center NUST MISIS.

The code is posted in the public domain, any interested person will be able to install and test this system.