Our current research is supported in part by the NIMH Human Brain Project under grant 1R01-MH66991 (and previously by the National Science Foundation under grants IIS-9811478 and IRI-9309273). Most of our projects are described below; for more details, see publications on Visual Cortex Modeling. For related projects, see Self-Organization and Concept and Schema Learning.
Work with the RF-LISSOM model has shown that it can develop realistic cortical structures when presented with approximations of the visual environment. However, the brain already has significant structure at birth, so environmental inputs cannot account for all of this self-organization. This ongoing research project explores a surprisingly simple but very effective way that an organism's genome can specify detailed cortical structures by generating training patterns internally.
Visual perceptual grouping is a process of identifying constituents that together form a group. In this project, a self-organizing map of spiking neurons was developed to understand the neural mechanisms of perceptual grouping. Grouping events were represented by the degree of synchrony among neural populations in the model, and the connection weights were self-organized to capture statistical regularities in the input. The resulting connection structure was consistent with experimental findings, and the functional performance of the model matched human performance in contour integration tasks. Furthermore, the same model was able to account for segmentation of multiple contours and contour filling-in. Altered input distribution caused the structural properties to differ among different areas in the model cortex, and as a result, functional performance also differed. Such a result provides a possible developmental explanation for similar psychophysical results.
Our goal is to understand how orientation tuning and direction selectivity simultaneously develop in the visual cortex. With this goal in mind we first built a SOM-based model that self-organized to represent these features. We then created a more low-level model, based on a delay adaptation learning rule that adjusts delays in a network based on spike timing. An early result shows that using the delay learning rule, the network develops into a map of directional selectivity. We have since built combined maps of directional selectivity and orientation in LISSOM, which shows how lateral connections can organize synergetically with the map.
Densely-connected self-organizing models of the cortex can be quite computationally intensive to simulate. We are working on two methods for making such simulations more practical. First, we have derived a set of scaling equations that allows small networks to be used as approximations for larger ones, while allowing the same parameters to be used for full-scale simulations once the concept has been demonstrated. Second, we are investigating how these scaling equations can be applied to a network as it is organizing, in order to develop a large, detailed final network in much less time (and using much less memory) than would otherwise be required. This growing laterally-interconnected self-organizing map algorithm is based on RF-LISSOM and is called GLISSOM.
SLISSOM is an extension of the LISSOM (or RF-LISSOM) model where the standard firing-rate neurons have been replaced by spiking neurons with leaky integrator synapses. The SLISSOM network self-organizes like the others but it can also represent segmentation and binding through synchronization and desynchronization of neuronal activity.
If self-organizing processes continue operating on the adult RF-LISSOM structure, tilt aftereffects can be modeled as a result of adapting lateral inhibition and re-normalization. Simulations show that the model can account for both the direct (small angle) and indirect (large angle) effects, while providing a clear computational role for both. This represents an important link between conscious perception and low-level, short-term synaptic plasticity.
In conjunction with research on the tilt aftereffect (above), a number of visualizations of orientation perception have been developed. These demonstrate how perception may be occurring in the adult cortex, with a level of detail not possible to achieve with current biological and psychophysical measurement methods.
In RF-LISSOM, the neurons receive inputs from local receptive fields on the retina instead of the entire retina as in SOM and LISSOM. This extension leads to realistic modeling of how the receptive fields develop into orientation, ocularity, and size detectors, how the neurons become globally ordered into orientation, ocular dominance, and size columns, and how patterned lateral connections develop between them. Such a structure forms a sparse, redundancy-reduced representation of the visual input and can account for several plasticity phenomena in the adult cortex. The self-organization is activity-dependent and driven by the input from the external environment, or by patterns generated internally.