22.07.2019

Towards Truly Intelligent Artificial Systems

Artificial Intelligence is currently a term that is used in various contexts and various domains. Even simple devices, like washing machines, are supposed to contain algorithms that contain AI. Despite quite phenomenal successes over the last years – yielding human-superior performance in image recognition tasks, beating the world’s best Go players, playing simple computer games and even the strategy game StarCraft better than humans, generating translations that are very useful and partially better than traditional computer translation approaches – all these systems optimize objective functions that are to maximize classification accuracy, that is, classifying image content or generating text that matches given annotations, or reward-based objective functions, such as staying alive as long as possible or winning a computer game. As a result, all these algorithms essentially apply complex statistical analyses to big data corpora, where the data is self-generated in the case of (computer) games, and exploit the detected regularities in the data. The generation of flexible, adaptive, and innovative behavior beyond the trained task, however, remains very hard to achieve. 

The Cognitive Modeling Group under the supervision of Prof. Martin V. Butz (and in collaboration with the MPI of Intelligent Systems in Tübingen as well as colleagues from the University of Otago, Dunedin, New Zealand) addresses this challenge by developing neuro-cognitive artificial intelligent systems. While these systems are still based on state of the art machine learning techniques, such as deep artificial neural networks, they focus on developing predictive, generative models of their environment rather than optimizing classification or maximizing reward directly.

Two journal papers, which were just published in the renowned Neural Networks journal as well as in the IEEE Transactions on Cognitive and Developmental Systems journal, underline the potential of their approach. It is shown that their systems can develop conceptual abstractions from sensorimotor experiences: the learning systems explore their behavioral capacities in complex, simulated environments and thereby develop abstract conceptual and contextual encodings from analyzing the encountered sensorimotor dynamics. That is, as we humans clearly experience fundamental different dynamics while, for example, walking, standing still, sitting down, lying in bed, grasping an object, transporting an object, or drinking out of an object (bottle, mug, etc.), these systems detect these differences and compactly store the experiences in, so-called, event-predictive encodings. Once such encodings have been learned, the developing learning systems are able to exploit these encodings for the generation of complex, hierarchical goal-directed behavior. As a result, the systems are able to plan on a much further reaching, conceptualized (event-predictive) temporal horizon, and to adapt their behavior on the fly to the current circumstances in a much more versatile manner.

While further research is needed to show that these systems can scale up to even more complex problems and environments, the capacity for abstraction and conceptualization does not only mimic human cognition and behavior but it may have the potential to lift current artificial intelligence techniques onto a new level beyond classification towards actually understanding and reasoning about the encountered problem domain. Clearly, once this is achieved, such systems may actually solve problems on their own, including some that the AI designers may have not thought about originally. 

Martin V. Butz

Theoretical Cognitive Modeling Background:  

Butz, M. V. (2016). Towards a Unified Sub-Symbolic Computational Theory of Cognition. Frontiers in Psychology, 7 (925). doi:10.3389/fpsyg.2016.00925

Two journal publiclations mentioned: 

Butz, M. V., Bilkey, D., Humaidan, D., Knott, A., & Otte, S. (2019). Learning, planning, and control in a monolithic neural event inference architecture. Neural Networks, 117, 135–144. doi:10.1016/j.neunet.2019.05.001
Gumbsch, C., Butz, M. V., & Marious, G. (2019). Autonomous identification and goal-directed invocation of event-predictive behavioral primitives. IEEE Transactions on Cognitive and Developmental System. 

Prof. Dr. Martin V. Butz
University of Tübingen
Computer Science, Cognitive Modeling
 +49 7071 29-70429 
martin.butzspam prevention@uni-tuebingen.de 

http://cm.inf.uni-tuebingen.de/