ERC Grant for AI-project led by TU Professor Constantin Rothkopf

Human information processing

Since the mid 19th century, the behavioral sciences have used laboratory experiments to reveal numerous systematic illusions, errors and biases in human perception, decision making, thinking and acting. But how can these phenomena be explained scientifically? Cognitive science, the twin science of AI, employs computational models to better understand human behavior.

The idea is that, if cognition is based on information processing, it has to adhere to the fundamental laws of computation. Indeed, the last decades have seen substantial progress in modeling, predicting and explaining human behavior including illusions, errors and biases through information processing algorithms.

Models of human behavior in the real world

While most behavioral experiments under laboratory conditions use short, highly controlled tasks with unambiguous goals, real world tasks such as navigating in an unfamiliar environment or even making a sandwich are characterized by a multitude of ambiguities, uncertainties and variabilities. As an example, although we are not aware of it, our eyes generate images of very poor quality, much worse than a consumer camera. Nevertheless, we humans are unmatched in the recognition and understanding of visual input.

Of course, that comes at the price of approximately 60 per cent of our brain’s computational resources being involved in the information processing that makes us see. “It’s no coincidence that current computer algorithms excel at clearly defined tasks with few uncertainties such as playing chess and go, but they are still hopelessly lost in everyday tasks governed by the full breadth of the world’s uncertainties: it’s because of the fundamental laws of information processing.” says Rothkopf.

Perspectives for cognitive science and AI

This is where ACTOR starts by considering not only classic experiments from perceptual psychology but also everyday tasks in the real world such as navigating unknown environments or even preparing a sandwich. “Based on computational models of sequentially perceiving and acting, to which we have contributed together with our colleagues for several years, we will develop a better understanding of human conscious and unconscious decision processes in natural tasks.” says Rothkopf. The results will not only further the understanding of human perception, cognition and behavior but are expected to also contribute to applications in which humans interact with adaptive systems such as in the area of human computer interaction or interacting with learning robotic systems.