Bases and implications of learning in the cerebellum — adaptive control and internal model mechanism
Introduction
Investigations of the cerebellum have been greatly advanced during the past century, and several lines of basic knowledge accumulated thus, to provide the basis for understanding the neuronal mechanisms and functional roles of the cerebellum. First, the neuroanatomical studies revealed that the cerebellum has a unique geometrical architecture represented by the checkered board map shown in Fig. 1. It is divided into the middle longitudinal area, vermis, and the intermediate and lateral parts of the hemisphere, which are further subdivided into 13 major longitudinal zones (A in the middle and B, C1, C2, C3, D1 and D2 on the right and left sides) (Groenewegen et al., 1979). At the vermis, grooves running along the right–left axis divide the cerebellum into ten lobules I–X, which continue to the hemisphere as lobules HII–HX (Larsell and Yansen, 1972). While certain divisions are missing in the posterior part of the B, C1 and C3 zones, and there are additional zones such as x in the anterior vermis (Fig. 1), the cerebellum contains more than one hundred divisions.
Second, the cerebellar divisions have uniformly structured neuronal circuits and hence are presumed to operate by the same local circuit principles. The elaborate network structures of the cerebellar cortex composed of the Purkinje, basket, stellate, Golgi and granule cells, and mossy and climbing fiber afferents have been analyzed extensively. More recently, Lugaro cells (see Lane and Axelrod, 2002), unipolar brush cells (see Dino et al., 1999) and serotonergic, noradrenergic, dopaminergic and acetylcholinergic afferents (see Schweighofer et al., 2004) have also been recognized as unique elements of the network. While various types of synaptic plasticity have been located in the network as possible memory elements (Hansel et al., 2001), the long-term depression (LTD) is unique being induced by the convergence of signals from both granule cell axons and climbing fibers to a Purkinje cell. The signal transduction underlying LTD has recently been analyzed extensively (see Ito, 2001, 2002). Earlier, the author proposed a cerebellar corticonuclear microcomplex (hereafter referred to as ‘microcomplex’) as a functional unit capable of learning with LTD as a major memory process (Ito, 1984). A microcomplex is composed of a microzone of the cerebellar cortex interconnected to a small group of cerebellar nuclear neurons (or vestibular nuclear neurons) and a small group of inferior olive neurons that project climbing fiber afferents. A microcomplex modifies the relationship between the input from mossy fiber afferents and the output from nuclear neurons in response to error signals conveyed by the climbing fibers.
Third, these microcomplexes are connected to different extracerebellar structures, and hence are involved in diverse functions. Since a microcomplex has an area of 10 mm2 in the cerebellar cortex, and since the entire cortex has an area of 50,000 mm2, a human cerebellum may contain 5000 microcomplexes (Ito, 1984). A prototype of a cerebellar control has been developed from the manner in which the vermis, including the A and B zones, and the intermediate part of the hemisphere including the C1–C3 zones are connected to the spinal cord and the brainstem. Figure 2 illustrates the involvement of the flocculus in the two-degrees-of-freedom control of the eye movements in response to the head movements and optokinetic stimulation. A relatively small number of specific functions have so far been localized in the cerebellar surface, as illustrated in Fig. 1. Locomotion is represented in the vermis of lobule V (Yanagihara and Udo, 1994) and the C1 and C3 zones of the rostral paramedian lobule (Apps and Lee, 1999), and the saccadic eye movements in the oculomotor vermis covering lobules VI–VIII (Noda and Fujikado, 1987; Barash et al., 1999). A major site involved in the eye-blink conditioning is the C1 and C3 zones of lobule HVI (Attwell et al., 2001). The scheme of adaptive control applies well to the evolutionarily old, medial part of the cerebellum connected to the brainstem and the spinal cord (Ito, 1984; Barlow, 2002). In particular, the vestibuloocular reflex adaptation (Nagao and Ito, 1991; De Zeeuw et al., 1998), eyeblink conditioning (Bao et al., 1998) and locomotion (Ichise et al., 2000) are now widely used as model systems for testing the cerebellar learning function under various pharmacological and genetic manipulations of the cerebellar neuronal circuits.
Fourth, during the course of evolution in vertebrates beyond birds toward humans, the cerebellar hemisphere expanded markedly in association with the development of the cerebral cortex. The cerebellar hemisphere is thus involved in voluntary movements such as pointing with a hand (Kitazawa et al., 1998), visuomotor tracking with a computer mouse (Imamizu et al., 2000) and the hand-grip control (Kawato et al., 2003) (Fig. 1). Furthermore, the posterolateral part of the hemisphere represents language (Petersen et al., 1989; Fiez et al., 1992; Gebhart et al., 2002). In this chapter, the author presents the effort to expand the adaptive control system concepts thus far developed in studies of the evolutionarily old, medial parts of the cerebellum involved in the brainstem and spinal cord functions to the evolutionarily newer, lateral parts of the cerebellum involved in the cerebral cortical functions. To this effort, the internal model mechanism is instrumental. It is still largely hypothetical, but it receives increasing support from neuroanatomy and imaging analyses of human brains and computational studies in robotics. The internal model mechanism explains a number of seemingly unexplainable observations in movements as well as mental actions, covering a broad range of central nervous system functions, and its verification will be a major target for future cerebellar research.
Section snippets
Forward model for voluntary movement
The intermediate part of the hemisphere including the C1–C3 zones, is connected not only to the brainstem and the spinal cord via the projection from the interpositus nucleus to the red nucleus and other nuclei, but also to the cerebral primary motor cortex (hereafter referred to as the motor cortex) via the projection from the interpositus nucleus to the thalamus. Since the motor cortex projects to the cerebellum via certain brainstem nuclei (pontine nuclei, nucleus reticularis tegmenti pontis
Sensory cancellation by prediction
In the above-mentioned mode of a forward model control, internal feedback replaces external feedback and makes the system perform effectively even if the external feedback is unavailable (c replaces b in Fig. 4). In another mode, internal feedback predicts a sensory consequence of a movement and acts to cancel it, which otherwise induces a sensation that disrupts the execution of the movement (c cancels b in Fig. 4). A clear example has been observed in fish cerebellum-like tissues, which
Automatization by the inverse model
After a long practice in motor learning, we become so skillful that we no longer need to be conscious about an executed movement. It often happens that, after a grueling match, Sumo wrestlers remember very little about how they have performed, which suggests that they performed largely automatically. Such skill automatization can be explained, at least in part, based on the forward model (Fig. 4), which enables us to perform a precise movement without paying attention to the consequences of an
Source of instruction signals for a voluntary movement
To complete the control system scheme shown in Fig. 4, Fig. 6, sources should be defined for instruction signals to the motor cortex and the cerebellum, which designate the content of the movement to be performed, for example, a desired trajectory of an arm movement. These instruction signals should come from the cortical areas devoted to motor planning and preparation, directly from the anterior cingulate gyrus (ACG), supplementary motor area (SMA) and premotor area (PMA), and indirectly from
Internal model for mental action
The most lateral part of the hemisphere (D1 and D2 zones) is connected to the cerebral association cortex. In humans these zones expand markedly and have been suggested to be involved not only in motor but also in certain cognitive functions such as language (Leiner et al., 1986). In contrast to the sensory and the motor cortices that receive stimuli from the external environment and send outputs to the external environment, the association cortex contains a self-sufficient internal loop in
Internal models for motor actions
Hitherto, we have discussed the possible mechanisms of voluntary movements and thoughts, separately. However, various motor actions that we daily perform integrate both movements and thoughts. One expects that such a motor action involves not only an internal model of a motor apparatus in the cerebellum (Fig. 4, Fig. 6) but also a movement-related mental model in the association cortex and its internal model in the cerebellum (Figs. 7 and 8). A brain imaging study demonstrated the
Perspectives
The internal model hypothesis of the cerebellum is based on a combination of three lines of approaches to solving problems pertaining the brain and mind: (1) neuroscience of neuronal circuits including synaptic plasticity, (2) system neuroscience based on modern control theories, and (3) mental model in psychology. It broadly covers problems of cognition and conation and, in particular, the mechanisms of physical and mental exercises that automate our behavior. It does not directly cover
Conclusions
The cerebellum has a fine compartment structure, which represents various discrete functions. Evolutionarily old, medial parts of the cerebellum is involved in the adaptive control of various brainstem and spinal cord functions, and long-term depression plays a key role in this adaptive control. To extend these views to evolutionarily newer, lateral parts of the cerebellum involved in cerebral cortical functions such as voluntary movement, perception and language, the internal model hypothesis
References (67)
Cellular and network properties in the functioning in the nervous system; from central patter generators to cognition
Brain Res. Rev.
(2003)- et al.
Delusions of alien control in the normal brain
Neuropsychologia
(2003) - et al.
Expression of a protein kinase C inhibitor in Purkinje cells blocks cerebellar LTD and adaptation of the vestibulo-ocular reflex
Neuron
(1998) - et al.
The control of forelimb movements by intermediate cerebellum
Prog. Brain Res.
(1997) - et al.
From will to action: sequential cerebellar contributions to voluntary movement
Neuroimage
(2003) Movement and thought: identical control mechanisms by the cerebellum
Trends Neurosci.
(1993)Cerebellar learning in the vestibulo-ocular reflex
Trends Cogn. Sci.
(1998)- et al.
Progr
Brain Res.
(2003) - et al.
Molecular analysis of nicotinic receptor expression in autism
Mol. Brain Res.
(2004) Memory traces unbound
Trends Neurosci.
(2003)