Researchers from HSE University in Nizhny Novgorod, MISIS and the Artificial Intelligence Research Institute (AIRI) have developed an algorithm that selects the best available neural network for facial recognition, taking into account the features of a mobile device. This new approach accelerates the selection of the most suitable neural network and allows the identification of people with an accuracy rate of up to 99%. The study was published in the IEEE Access journal. The source code is available on GitHub.
This research received a grant for research centres in the field of artificial intelligence provided by the Analytical Center for the Government of the Russian Federation.
Neural networks that can recognise faces usually require a large amount of computational power to work. There is no one perfect network for every device, since their features differ significantly: on one smartphone, one particular neural network can recognise faces quickly, while on another it will work with an unacceptable delay for the user.
There is a lottery ticket hypothesis in machine learning. It states that in a very deep network, you can ‘win the lottery’ — select one part of the neurons and the connections between them so that the accuracy of the resulting subnetworks turns out to be almost the same as that of the original network. There are a huge number of such subnetworks, and it is extremely difficult to find the most accurate one that would recognise a face in a given time. You can measure accuracy using a special face datasets by going through several randomly selected subnetworks, and thus eventually find the best one. However, this selection process may take weeks, and only a small part of the possible subnetworks will be checked.
Researchers from the HSE Laboratory of Algorithms and Technologies for Networks Analysis (Nizhny Novgorod), MISIS and the Artificial Intelligence Research Institute (AIRI) proposed ‘personalising’ the neural network selection system for facial recognition. It takes into account the features of a particular mobile device and allows you to make the selection process as fast as possible — usually only taking 5-10 minutes. The researchers proposed using a comparator, an algorithm that, without taking measurements, selects the most accurate of two proposed networks in a second until only the most suitable one remains.
Let's say you have 500 small subnetworks. All of them are sorted, and a hundred of the best, expected to be the most accurate, stand out from them. Then, with the help of ‘mutation’ and ‘crossing’, new, even more accurate subnetworks, are generated from the selected ones. Then the process repeats. During these ‘mutations’, some parts of the symbolic description of the subnetwork are randomly changed; during ‘crossing’, half of the description of one subnetwork is added to the half of the second one. This approach is called a genetic algorithm, or an evolutionary search. To select the top solutions, we proposed using a gradient boosting comparator, a popular machine learning algorithm for classifying tabular data. It turned out that it works more accurately and better and learns much more effectively.
Experiments have shown that the proposed method is resistant to various facial positions. The algorithm allows you to find models where identification accuracy reaches 97-99%.
The authors of the technology have also developed a demo mobile application for devices running on the Android operating system. It allows you to analyse the technical capabilities of the gadget, select a suitable neural network, measure its operation time and find the same person in two photos selected from the device's gallery. The development code for this has been published on an open access platform.
IQ
Andrey Savchenko
Professor, HSE Campus in Nizhny Novgorod