ABSTRACT
This chapter introduces a theory of how a neural network capable of learning and generating a temporally adaptive conditioned response (CR), the conditioned nictitating membrane response (NMR) of the rabbit, might be implemented in brainstem and cerebellar circuits underlying this behavior (Moore & Berthier, 1987; Thompson, 1986; Yeo, 1987). We characterize the conditioned NMR as temporally adaptive because its topographical features, latency and form, depend on the timing of the unconditioned stimulus (US) in relation to the conditioned stimulus (CS)(Gormezano, Kehoe, & Marshall, 1983). Unlike simpler CRs, such as those evinced by some invertebrate preparations, a CS for a temporally adaptive CR such as the NMR does not trigger an invariant reflexive response of fixed latency and form. Instead, it sets an occasion for a variable response that is sensitive to the temporal dimension of the task. A temporally adaptive CR can be regarded as a skill that requires a subtle resolution of forces (Desmond, 1988; Desmond, in press; Desmond & Moore, 1988; Moore, Desmond, & Berthier, 1989). A number of authors have previously stressed this point (Kimmel, 1965; Levey & Martin, 1968; Martin & Levey, 1965). The temporally adaptive features of the conditioned NMR encompassed by our model are listed here:
The latency of the CR with respect to onset of a conditioned stimulus (CS) changes during training. Initially, the CR appears as an enhanced unconditioned response (UR). The nascent CR can also be detected with CS-alone probes at this stage. It next appears just prior to the onset of the unconditioned stimulus (US). At this stage of early CR acquisition, the CR anticipates the UR and tends to blend with it to assume a smooth unimodal form. With further training the CR emerges as a shadow of the UR cast forward in time toward CS onset. In this regard its development is analogous to the growth of a goal gradient in which the vigor of a sequence of goal-oriented actions progressively increases backward from the locus of reinforcement toward the initial segments of the sequence.
The latency of the CR depends on the CS–US interval employed in training. If this interval is long the CR is delayed (Pavlovian inhibition of delay). Furthermore, the amplitude of the CR grows progressively (ramps) toward a peak at the temporal locus of the US, as can be demonstrated with CS-alone probes. If the temporal locus of the US changes (e.g., with the introduction of a new CS–US interval), the peak amplitude of the CR changes progressively toward the new US locus.
In trace conditioning protocols, in which CS offset precedes the US, CR initiation and peak amplitude tend to occur within the trace portion of the CS–US interval (i.e., at the same temporal locus as in forward-delay conditioning protocols).
