Cognitive Pilot has introduced a DVSE approach to train neural networks when creating drones

2021: Implementation of the DVSE approach for selecting data from a video stream when training neural networks

On March 17, 2021, Sberbank announced that Cognitive Pilot, a developer of artificial intelligence systems for unmanned vehicles, is part of the Sber ecosystem, has introduced the DVSE (Deep Visual Semantic Embeding) approach to select diverse, representative data from the video stream when training neural networks. The process is based on the use of artificial intelligence and does not require human participation.

In order for a neural network to recognize objects of a road, field or other environment with industrial accuracy and ensure the safety of driving at any time of the day, it must be trained on large amounts of data. Manually selecting this data from a video stream with a duration of more than a billion frames is very long and tedious. The DVSE approach allows the neural network to automatically select diverse and representative data from the video stream using a metric to compare different frames. The mechanism also allows you to filter out excess data from the video stream that does not affect the learning process.

The implementation of the approach solved a difficult task that saves tens of man-years in the creation of autopilots and greatly simplifies the development process in this area. If earlier it took years to process the video stream when training neural networks and creating datasets, now just press the button.

{{quote 'author = said Alexander Vedyakhin, First Deputy Chairman of the Management Board of Sberbank | Our colleagues at Cognitive Pilot have demonstrated a huge difference between artificial intelligence systems and traditional approaches. This development reduces development time many times and radically simplifies the creation of smart solutions in one of the most sought-after sectors of the market - drones. Thanks to this technology, we will be able to increase the functionality of existing solutions for autonomous management of agricultural and rail transport and scale our projects to Russian and foreign markets,}}

{{quote 'author = noted Olga Uskova, CEO of Cognitive Pilot |

We have extensive experience in training neural networks and creating datasets. Unlike many players in the artificial intelligence market for unmanned vehicles, who work mainly with ready-made data, public datasets and use them in applications for one direction (either agro or automotive, or otherwise), we have accumulated solid expertise in training neural networks on our own datasets in several segments at once: agro, rail and auto,}}