Previous Up Next
Recurrent ANN models, so called ``attractor models'', can be regarded as mathematical instantiations of Hebb's cell assembly theory of cortical associative memory (Hebb D. O. The Organization of Behavior, John Wiley, 1949). To evaluate the biological relevance of these abstract models we have investigated to what extent a network comprised of biological neurons is capable of operating as a content addressable memory (CAM) in much the same way as e.g. a Hopfield network. In particular, we have studied such a network's ability to display after-activity as proposed by Hebb, and to perform pattern completion and reconstruction.
More recently, properties of the assembly have been studied in detail. Experiments show that real neurons don't work in their upper region of their frequency range. Biologically acceptable mechanisms to obtain this type of activity have been studied. The use of saturating synapses was found to give low frequencies in the isolated assembly (FransÚn and Lansner 1995) as well as in a network of overlapping assemblies (FransÚn and Lansner 1994), (Lansner and FransÚn 1994). In an extended columnar model further discussed below, low frequencies were also obtained.
The operation of the isolated assembly has also been compared to experiments and simulations of the hippocampus, a part of the brain intimately related to memory. Simulations varying both synaptic strength and neuromodulator application corresponded well to reported results.
Currently, we investigate effects when going in the direction of artificial networks by adding more features from the corresponding ANN. We examine the effect on pattern completion and reconstruction capabilities when the bias values (of the Bayesian learning rule used to derive the network connectivity) are introduced. The overall relation of attractor ANN:s to models of associative memory in biology and psychology was reviewed in (FransÚn and Lansner 1995).
An extended assembly model with increased realism in connectivity has been developed. It is composed of cortical columns allowing the connectivity to be both locally dense and globally sparse and non-symmetric on the cell-to-cell level. In the figure the connectivity is displayed (left to right) on the column-, the cell-, and the cell compartment level. The standard assembly operations were retained in this extended model. Properties of the column, specifically its after-activity abilities, have been studied in detail.