A Scalable Model of Cerebellar Adaptive Timing and Sequencing: The Recurrent Slide and Latch (RSL) Model

摘要

From the dawn of modern neural network theory, the mammalian cerebellum has been a favored object of mathematical modeling studies. Early studies focused on the fanout, convergence, thresholding, and learned weighting of perceptual-motor signals within the cerebellar cortex. This led in the proposals of Albus (Mathematical Biosciences, vol. 10, pp. 25–61, 1971; Journal of Dynamic Systems, Measurement, and Control, vol. 97, pp. 220–227, 1975) and Marr (Journal of Physiology (London), vol. 202, no. 2, pp. 437–470, 1969) to the still viable idea that the granule cell stage in the cerebellar cortex performs a sparse expansive recoding of the time-varying input vector. This recoding reveals and emphasizes combinations (of input state variables) in a distributed representation that serves as a basis for the learned, state-dependent control actions engendered by cerebellar outputs to movement related centers. Although well-grounded as such, this perspective seriously underestimates the intelligence of the cerebellar cortex. Context and state information arises asynchronously due to the heterogeneity of sources that contribute signals to compose the cerebellar input vector. These sources include radically different sensory systems—vision, kinesthesia, touch, balance and audition—as well as many stages of the motor output channel. To make optimal use of available signals, the cerebellum must be able to sift the evolving state representation for the most reliable predictors of the need for control actions, and to use those predictors even if they appear only transiently and well in advance of the optimal time for initiating the control action. Such a cerebellar adaptive timing competence has recently been experimentally verified (Perrett, Ruiz, and Mauk, Journal of Neuroscience, vol. 13, no. 4, pp. 1708–1718, 1993). This paper proposes a modification to prior, population, models for cerebellar adaptive timing and sequencing. Since it replaces a population with a single element, the proposed Recurrent Slide and Latch (RSL) model is in one sense maximally efficient, and therefore optimal from the perspective of scalability.