Tech Talk. Technology of the future

Computer vision, deep learning and artificial intelligence (AI): Dr. rer. nat. Julian Schweizer is a Data Scientist at Nect responsible for the Robo-Ident technology of today and tomorrow. As part of his dissertation at DESY in Hamburg, he conducted research at the boundary between particle physics, cosmology and string theory and then moved into the field of Data Science. Today, he uses his know-how to ensure that our technology gets better every day and that we can maintain our position as an innovation leader. Julian Schweizer talks about what computer vision and deep learning even mean and how this field can change the future of Nect.

What does "deep learning" mean?

Experts in computer vision, i.e. the computer-aided understanding of image content, have been concerned with Deep Learning at least since its potential for image recognition became clear in 2012. But what does "deep learning" actually mean? In Deep Learning, human-made concepts such as edges and corners, which were previously considered essential for automatic image understanding, take a back seat; instead, one lets the computer itself calculate which elements of an image are relevant in order to draw a certain piece of information from it. Today, Deep Learning and artificial intelligence are almost synonymous.

A virtuous cycle of research and application of this technology has revolutionized a whole range of fields of work that process unstructured data such as images or text - for example, computer vision, recommender systems or natural language processing.

These fields have in turn inspired each other in various approaches. The most recent example is the application in computer vision of the so-called "transformer" models behind the huge advances in Natural Language Processing in recent years. An exciting aspect of this is that the computer decides not only which elements are relevant, but also which regions of the image - dynamically from image to image. If computational cost is not an issue, these models seem to be a promising next step to better deal with the extreme variety of images of what is actually the same content (think, for example, of all the ways to photograph zebras - the options are almost endless).

From the other side, there are developments from both the hardware and software sides to run more and more computer vision and AI software directly on user devices. Special chips to run certain types of AI models with minimal latency and power consumption are in almost every smartphone these days. These are exactly the ones behind the facial biometric logins and awesome photo apps that have been proliferating for some time.

Computer vision at Nect

At Nect, we're always keeping a close eye on the latest technological developments to see how they can help us in our mission - to revolutionize the world of online identification. "Transformers" in computer vision are highly predictive, and we are ready as soon as it becomes apparent that their potential could also be realized in our applications. This could, for example, allow us to deal even better with the different image backgrounds and lighting prevalent in our users' homes.

Being able to run more complex AI models directly on the end device also opens up exciting options for us. In particular, we can make the recording process in our app easier for our users. Sometimes they don't find this very easy. To help them better and in real time, we will increasingly bring computer vision and AI components into the app. We can then respond directly to what the user is doing and provide them with a smooth, interactive Nect experience. Already today, we are the most popular identity verification app in terms of user experience. We want to maintain and further expand this status. With Computer Vision, much more will be possible here in the future.