Small networks of empirically derived adaptive elements simulate some higher-order features of classical conditioning

-Previously, we developed a single-cell mathematical model of the sensoo' neurons in Aplysia (Gingrich & Bvrne, 1985, 1987 ). This single-cell model accurate@ simulated many aspects of empirical@ observed neuronal plasticity that contribute to simple forms of nonassociative and associative learning. In the present study, we incorporated this empirically derived adaptive element into small networks attd examined the ability Of these networks to simulate second-order conditioning and blocking. When the single-cell model was incorporated into a three-cell ;;etwork (Hawkins & Kandel, 1984), we fbltnd that cottstraints imposed by the empirical data limited the ability o['rtte network to simulate both second-order conditioning and blocki;~g. On the other hand, we found t/tat the detailed descriptions ~)[subcellalar processes unmasked phenotnena relevant to the simulation of blocking, that are not captured by h'ss detailed models. We also ineorl?orated the model of the sensory neuron into a lateral inhihition-O'pe network consisting o[ five ele,tettts. This network succes,silhlly simulated both second-order conditi(ming and blocking more readily thatt the three-cell network. Keywords--Ap@sia, Blocking, Lateral inhibition, Learning, Models. Neural networks, Second-order conditioning. Sensory neurons. 1. I N T R O D U C T I O N Computer simulations of neural networks have proven to be a valuable tool in helping to understand how assemblies of neurons may perform the wide spectrum of adaptive information processing observed in biological systems (e.g., Bienenstock, Cooper & Munro, 1982; Desmond & Moore, 1988; Fukushima, Miyake & Ito, 1983; Gelperin, Hopfield & Tank, 1985: Gelperin, Tank & Tesauro, 1989: Grossberg, 1988: Grossberg & Levine, 1987; Hopfield, 1982: Klopf, 1988: Pearson, Finkel & Edelman, 1987: Sejnowski & Rosenberg, 1986; Sutton & Barto, 1981). Many models of neural networks are composed of simple interconnected elements that take the weighted sum of their inputs and generate an Acknowledgements: We thank Drs. M. Mauk and A. Susswein for their comments on an earlier draft of the manuscript and Mr. S. Patel k)r assistance with computer programming and graphics. This research was supported by Air Force Office of Scientific Research Grant 87-0274 and National Institute of Mental Health Award K02 MH00649 and Fellowship F31 MH09895. Requests f~or reprints should be sent to Dean V. Buonomano, Department of Ncurobiology and Anatomy, University of Texas Medical School, P.O. Box 20708. Houston, TX 77225. output via an activation function. The weights between network elements change according to a "'learning rule." While this approach has been successful, a fundamental issue in neural network modelling, which has not been addressed adequately, is the level of detail necessary to model individual neurons and how such detail affects the global properties of networks. One way of addressing this issue is to examine the ability of neural networks consisting of elements based on the detailed properties of neurons to simulate a well-defined task for which empirical data are available. One such well-defined task is classical conditioning. The parametric features of classical conditioning have been examined extensively, and thus, there is a large body of data with which the performance of the network can be compared. Moreover, cellular mechanisms underlying some forms of associative plasticity have been identified, and thus, it is becoming possible to incorporate detailed descriptions of the cellular processes involved in neuronal plasticity into the individual elements of the neural networks. During first-order classical conditioning (e.g., Mackintosh, 1974: Parlor, 1927), an animal is pre-