My doctoral research focused on studying adaptive
behavior in computational vision systems based on coupled map lattices,
and interpreting them as models of medium and large-scale neuronal dynamics.
Each unit or site of the lattices is a bifurcating or chaotic element, with
local recurrent connections and a variety of possible short and/or long
range coupling functions with other such units. Such systems
may also be considered discrete time cellular neural networks, when strictly
short-range coupling is used. I have employed these systems
to model psychophysical phenomena and perform computer vision tasks, including
3-D object recognition. Information theoretic principles are used to analyze
network behavior and guide the evolution of networks.
This work, and related applications of synchronization in oscillating networks, has many potential applications in search of images and audio signals. The general principles may also be merged with symbolic methods and representations, and are applicable to the processing and study of linguistic communication and dialog processes.
I have also been involved with research in human interfaces for complex design tasks, algorithmic composition methods in music and visual art, and with visualization and simulation methods in general.
Previous
work has demonstrated
1) The potential of forming spatial representations by the spatial distributions of oscillation clusters; spatial patterns of high contrast, through spatial self organization processes, control the clustering. A dynamical interpretation of attention as local increase in synchrony - which is projected as an increased coupling effect to other regions - is proposed.
2) Partial synchronization processes in the transient evolution of spatially regular systems can support recognition of 3-D objects. This can be consider as recognition by forming dynamical recognizers, which implicitly form a representation space by locating the "decision states" at particular points in the space formed by partition cells of the dynamics. This allows the representations to support metric similarity functions.
3) Such systems, due to rich dynamical structure, may offer advantages in ease of learning representations and useful computations. Genetic algorithms have been used to construct dynamical recognizers for the image families for a three dimensional object. Good performance on this task is achieved with very low evaluations (3000). The task is difficult for humans, indicating that solutions to the representation problem are dense in the parameter space. Further work comparing the ease of learning in such networks relative to other sigmoidal unit networks (by GA or other methods) would be interesting.
These previous research directions are only a beginning effort in intteligent computation with spatio-temporal chaos, and deserve further study and development. Such extensions have been documented in the "future work" sections of relevant publications.
In addition, future related projects of interest include:
1) The use of modulated coupling to represent behaviors, where the behaviors are sequences of visited subspaces in the dynamics. The interaction of emotions and behavior can be explored in this framework. Historical ideas of competitive control through inhibition may be supplemented by (possibly) more realistic models involving competition of synchronization or clustering processes under emotional modulation, where emotions are conceived of as points in a control space of coupling and bifurcation parameters. These two dimensions may correspond roughly to arousal state and mood.
2) The refinement of evolutionary computing strategies to explore such systems, including the embedding of such systems as "animats" in artificial environments.
3) Learning and recognition networks which do not employ evolutionary computing directly (e.g. the creation of dynamical recognizers for object families) but rather to construct topologies, coupling types, baselines and ranges for task specific behavioral control, with more biologically plausible (resembling Hebbian and anti-Hebbian learning) dynamics.