Aug 02 2007
CogSci 2007 Day 2
Day 2
A fairly good day at the conference today, though I must admit that I still think the conference is a bit too dominated by people with Psychology backgrounds, and not enough representations from the other disciplines. Cognitive Science has always been hailed for its highly interdisciplinary nature, but there is a dearth of other disciplines lately (AI, neuroscience, etc) at the conference. Though there are new prizes this year for the best computational models of various cognitive processes or areas, a good development. But I find things exemplified by people like the morning Plenary speaker Walter Kinstch (Latent Semantic Analysis) or the evening Plenary speaker John Anderson and his ACT-R cognitive models (a symbolic / expert system type of approach, but with low level mechanisms to model and account for timing and other empirical brain data).
Anyway, there was one particular find (for me) today, a paper by Anthony Morse and Tom Zimke from the University of Skovde on cognitive robotics and enactive perception. A bit tough to get the idea across in a few sentences, but they are using so called reservoir systems (such as Liquid State Machines or Echo Nets) as a way to set up complex dynamic that can be perturbed by incoming (an internal) stimuli. The basic idea is that the combination of the complex dynamics of the system, perturbed by the inputs actually simplifies relations in the sensory/motor associations, and makes them easier to identify. They use the analogy of dropping a stone into a pond of water, the ripples afterwords represent the complex dynamics of the perturbing event. This is really quite similar to Freeman’s (and me and Kozma’s) viewpoint of how the K-models dynamics function in the olfactory system to enable active perception and categorization of smells (and associated behavior). Freeman and I have often used the metaphor of snowflake formation, in the right conditions the atmosphere is in a highly unstable state far from equilibrium. The introduction of an perturbance (a dust particle for example) causes a reaction to catalyze e.g. the snowflake crystalizes. This crystalization is a kind of complex dynamics that encodes the original perturbance in a new way.
The K-models share similar properties with such reservoir systems. They are sparse randomly interconnected networks of artificial neurons. The networks create a natural, very high-dimensional dynamical state-space. The high-dimensional state space can have different inputs (and input sequences) cast into them, and the dynamics react in different ways to the inputs. Unfortunately (for me), I think Morse is way ahead of demonstrating this type of dynamics in a complete cognitive agent, and how it might work to yield active, situated, embodied perception.