This article reviews models of the cerebellum and motor learning, from the landmark papers by Marr and Albus through those of the present time. The unique architecture of the cerebellar cortex is ideally suited for pattern recognition, but how is pattern recognition incorporated into motor control and learning systems? The present analysis begins with a discussion of exactly what the cerebellar cortex needs to regulate through its anatomically defined projections to premotor networks. Next, we examine various models showing how the (...) microcircuitry in the cerebellar cortex could be used to achieve its regulatory functions. Having thus defined what it is that Purkinje cells in the cerebellar cortex must learn, we then evaluate theories of motor learning. We examine current models of synaptic plasticity, credit assignment, and the generation of training information, indicating how they could function cooperatively to guide the processes of motor learning. (shrink)

The data on the cellular mechanism of LTD that is presented in four target articles is synthesized into a new model of Purkinje cell plasticity. This model attempts to address credit assignment problems that are crucial in learning systems. Intracellular signal transduction mechanisms may provide the mechanism for a 3-factor learning rule and a trace mechanism. The latter may permit delayed information about motor error to modify the prior synaptic events that caused the error. This model may help to focus (...) future cellular studies on issues that are particularly critical for a computationally viable concept of cerebellar plasticity, [CRÉPEL et al.; HOUK et al.; KANO; LINDEN; VINCENT]. (shrink)

Neural models are needed at many levels, ranging from the biophysically based to connectionist abstractions. The adjustable pattern generator (APG) model of the cerebellum discussed in the target article by Houk, Buckingham, and Barto operates at an intermediate level and is constrained by a large body of anatomical and physiological data. The associative memory model originally proposed by Marr (1969) and Albus (1971) on the basis of the microcircuitry of the cerebellar cortex is still relevant today, although it needs to (...) be coupled to the problem of motor pattern generation. In considering the problem of cerebellar learning, one needs some source of training information to adjust network performance so that it performs fruitful interactions with the environment; however, it is not true that this “teacher” needs to be particularly smart. We address these and other important issues raised by the commentators. (shrink)