Computer vision and visions of the future at Swisscom’s Digital Lab
On Thursday, 14 September 2017, Swisscom held its monthly Digital Lab meet-up event on the EPFL campus. Artificial intelligence research was the overarching theme at the Digital Lab this time around. Pierre Baqué, a doctoral candidate at the Computer Vision Laboratory at EPFL, was invited to speak at the event about computer vision, an exciting interdisciplinary field with much potential for the future.
Addressing an audience of corporate, scientists and researchers with limited knowledge of the discipline (for most of them), Pierre’s talk covered the broad strokes of computer vision and machine learning techniques and tied it in with his own research in artificial intelligence algorithms. Pierre explained that computer vision focuses on how to condition computers to identify specific objects in digital images or videos. According to the theory, an artificial system can be ‘trained’ to extract specific information from a digital image or video and then complete a task based on that data. He noted that technologies are already in use in areas as wide-ranging as surveillance, counter-terrorism and statistical analysis. Computer vision and machine learning can also be combined to create life-like objects, such as a human face. After applying an algorithm to map the details of the face, another algorithm is used to flesh out the basic 2D design and create a full 3D model that is able to fool the human eye from a distance.
This was followed by a look at the limitations of computer vision, namely the challenge of decoupling one object from another, or ‘segmentation’. For example, a computer may be able to recognise a person or a bicycle when viewed separately, but would encounter difficulties if the person was actually riding the bicycle in the image. Pierre went on to explain how a computer’s capacity to recognise specific images can be enhanced through the use of conditional random fields (CRF) and generative adversarial networks (GAN), where two neural networks are configured to compete with one another to create synthetic data which is similar to the known input data. This enables full, convincing images to be created.
Pierre concluded by stating that even though GAN is not yet able to solve standard AI tasks outside the realm of image generation, he expected more research on this topic to be published in the coming years. In that context, Swisscom intends to work with Master’s students at EPFL on the application of different technologies for the benefit of its customers and the industry as a whole.
For more information about Swisscom Digital Lab and to find out about the initiative’s upcoming events, please contact firstname.lastname@example.org and/or register to our meetup group https://www.meetup.com/Swisscom-Digital-Lab/.
About Digital Lab
The result of a strategic partnership between Swisscom and EPFL, Digital Lab is an initiative whose aim is to realize applied research projects with the various labs at the EPFL, to share the learning in various fields, foster collaboration across disciplines and institutions, and set out aspirations for the future. Launched just under 15 months ago, Digital Lab has built a meet-up community (one of the largest meetup in Western Switzerland) of over 1,000 innovators, scientists and researchers specialising in digitisation, AI, state-of-the-art technologies and other related fields. The Lab itself is located on the EPFL campus in Lausanne, Switzerland.