The difficulty in developing artificial intelligence is not so much transferring complex calculations that represent major intellectual challenges for humans to computers and machines. The far greater challenge is to teach computers the skill of learning from experience that characterizes humans. Tasks that a human finds simple can quickly bring an AI system to its limits. Human abilities such as intuitive action, social and emotional intelligence, and the use of multiple sensory impressions to create one single image, cannot be described using formal mathematical rules.
Facial recognition is a good illustration of this problem. Human facial expressions are primarily determined by eight of the total 26 facial muscles used, among other things, to express our feelings and support non-verbal communication. Humans not only recognize familiar faces intuitively but also learn how to read these faces quickly to identify feelings. In contrast, a machine being used for facial recognition will identify a laughing face as a diff erent person. The identification and classification of feelings is an even bigger challenge. Long periods of training using large datasets and many images are required before algorithms can reliably recognize and classify faces. To do so, the systems use a machine-based learning procedure known as »deep learning«. It is based on multilayer neuronal networks whose organization was inspired by structures in the human brain. The procedure revised parameters in the learning phase for as long as it takes to be able to reliably recognize faces. The goal is to achieve the same quality as human information processing and, in some cases, to even exceed it.
Self-learning systems have a high potential for various fi elds of application. Using deep learning, systems can acquire knowledge; filter relevant observations out of large data volumes; draw logical conclusions from them; and – as impressively demonstrated by the thousands of years-old game ›go‹ – even develop their own action strategies.
Deep learning is increasingly being used in security-critical application areas, such as autonomous driving, medical applications or distributed power systems. The fulfilment of security-relevant parameters is essential for successful use in the relevant areas.
The Deep Learning competence cluster gathers all the opportunities and risks in the fi elds of deep learning, machine learning, and artificial intelligence and bundles OFFIS’ competences into a single crossdisciplinary research strategy.