Institute for Theoretical Physics
Goethe University Frankfurt/Germany
Daily research in artificial intelligence (AI), robotics and computational neuroscience involves the solution and simulation of concrete and specific cognitive tasks, like locomotive control, visual object recognition and so on. Considering the long-term perspective of an autonomous and selfmotivated artificial cognitive entity, possibly with humanlevel cognitive capabilities, the quite fundamental question then arises whether top-down or bottom-up approaches will allow us to achieve this goal.
Within the bottom-up perspective, currently the mainstream approach, one concentrates on algorithms and architectures specifically engineered for the solutions of concrete cognitive problems, like speech or object recognition. This approach is highly successful to date, resulting in powerful industrial-level applications. It is however unclear whether the bottom-up approach is suitable for a comprehensive understanding of the human cognitive system as a whole and for the long-term goal of constructing artificial human-level and self-motivating AIs.
Within the top-down approach one considers on the other side specifications suitable for complete cognitive systems with initially only limited cognitive capabilities, but with the potential of evolving gradually in complexity. The top-down approach draw support from the fact that self-organization is centrally important both for the neural activity of the adult brain (von der Malsburg 1998), as well as for the global structural organization (Price and Willshaw 2000).
We examine here a first general question: A complete cognitive agent (Gros 2008) is not just a complex inputoutput device, driven by the sensory input, but characterized by autonomously generated internal activities (Gros 2005).
How and when do then these internal states and activity patterns build up correlation with the information it receives via the sensory data input stream? Based on the principle of competitive neural dynamics we propose that correlations between internal and external states are generated whenever the later are able to influence and modify the internal thought processes.
Performing simulation of the resulting generalized neural net (Gros 2007). we find that the system performs automatically a non-linear independent component analysis (Gros and Kaczor 2008). We have therefore obtained an example
Copyright c 2008, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
for an emergent cognitive capability, indicating that a topdown approach is feasible
As a second topic we examine the role of diffusive emotional control both for global homeostatic regulation and in the context of the motivational problem (Gros 2009). That standard approach within artificial intelligence is to develop algorithmic solutions for specific cognitive tasks, with the assumption that motivations and goal-setting for an eventual human-level cognitive agent could be added at a later state. We argue that learning and motivational control are intrinsically related for biological cognitive systems and that true human-level cognitive capabilities can be achieved in the long term only when motivational drives in the form of diffusive emotional control are integral parts of the evolving cognitive agent.
Gros, C., and Kaczor, G.
2008.
Learning in cognitive systems with autonomous dynamics.
Arxiv preprint arXiv:0804.1306
.
Gros, C. 2005. Self-sustained thought processes in a dense associative network. In Lecture notes in computer science , volume 3698, 366–379. Springer.
Gros, C. 2007. Neural networks with transient state dynamics.
New Journal of Physics 9:109.
Gros, C. 2008.
Complex and Adaptive Dynamical Systems .
Springer.
Gros, C. 2009. Emotions, diffusive emotional control and the motivational problem for autonomous cognitive systems. In Handbook of Research on Synthetic Emotions and
Sociable Robotics: New Applications in Affective Computing and Artificial Intelligence . J. Vallverdu, D. Casacuberta
(Eds.), IGI-Global. (in press).
Price, D. J., and Willshaw, D. J. 2000.
Mechanisms of cortical development . Oxford University Press.
von der Malsburg, C. 1998. Self-organization and the brain. In The handbook of brain theory and neural networks . MIT Press. 840–843.