17 Feb 2022
To round off the KI Lectures series, computer scientist Thomas Seidl shared insights into the challenges of machine learning – and showed how computers are becoming creative.
How smart can AI get? Rounding off the virtual KI Lectures series, Prof. Thomas Seidl, LMU computer scientist and one of the three directors of the Munich Center for Machine Learning (MCML), made this question a key part of his lecture. He illuminated various tasks of machine learning and described how computers accumulate and systematize experiences and infer rules from them. “A better understanding of relationships helps us make better predictions and take better actions,” says Seidl.
Machine learning is already being used in numerous domains and has achieved a lot in areas such as image and speech recognition. With regard to the intelligence or “cleverness” of AI, Seidl raised the question as to what we actually mean by this. An ability that is often associated with intelligence is creativity. Contrary to what some have claimed, a computer is very much capable of creative achievements, as Seidl explained using the example of the computer program AlphaGo. “What’s important for creativity is the play impulse, and if I introduce that into a computer, then something creative transpires,” says Seidl. It is vitally important, he explains, to consider what you want in advance. Ethics, for example, cannot be incorporated after the fact. “Fundamental things such as privacy by design or ethics by design must be factored in when designing networks.” Moreover, the quality of results also depends on the quality and availability of training data. But even if there is a lack of training data or there is no annotated training data, you do not have to throw in the towel; rather the machine can make optimum use of the data and recognize patterns with the aid of techniques such as machine learning with few labels. “AI is already capable of doing a lot, and the limits are not yet apparent,” says Seidl with conviction. What is important, he emphasizes, is to really think about what we want to achieve with AI and how to design its use.