Trends in Neurosciences
Parallel neural networks for learning sequential procedures
Section snippets
Previous concepts of motor control
Many neuroscientists have proposed hypothetical schemes for motor control10, 11, 12, 13, 14, three of which are shown in Fig. 1. They can be classified into two groups, one stressing serial information processing and the other stressing parallel information processing. The former can be seen in the conceptual model of Kawato et al.13 (Fig. 1A), which illustrates the sensorimotor processes that are minimally required for a simple movement, such as reaching. Here, the location of the target is
New concept of motor control and procedural learning
Our concept is illustrated in Fig. 2, in which we outline the process of learning a sequential procedure (Acts 1–3). Evidence supporting this concept is described in the following section and summarized in Boxes 1 and 2.
Initially, the serial sensorimotor process is executed in a discrete manner for each elementary action (vertical connections in Fig. 2A). As the subject repeats the actions in a fixed order, new connections are formed between the mechanisms for individual actions (Fig. 2B and C
Experimental evidence obtained from trial-and-error learning of a sequence
We used a sequential button-press task with trial-and-error processes (’2 × 5 task’ for monkeys and ’2 × 0 task’ for humans)15 (Fig. 3). Monkeys learned a set of sequences repeatedly until they could perform the sequences highly skilfully. This allowed us to study the neural mechanisms that are involved in memory storage and retrieval processes. The 2 × 5 task also allowed us to study the neural mechanisms that are involved in the learning of new procedures because we could always generate a
Neural correlates of the spatial and motor sequence mechanisms
Our scheme (Fig. 4B) is similar to that proposed by Allen and Tsukahara10, with subtle but important modifications: the connections between the cerebral cortical areas, and the basal ganglia and the cerebellum are now bi-directional, thus forming loop circuits. These loop circuits can be classified into two groups, one using spatial coordinates and the other using motor coordinates. In this article, we refer only to eye- or head-centered coordinates for spatial representation, although there
Communication between the parallel mechanisms
We have suggested that a sequential procedure is acquired independently by the two different sequence mechanisms (spatial sequence mechanism and motor sequence mechanism). However, in order to acquire and execute a sequence adequately and efficiently, these mechanisms must cooperate or compete with each other. In these processes, the premotor area and the pre-SMA might have important roles.
Motivational value is attached by the basal ganglia
Using evidence from recent studies57, 58, 59, 60, we propose that the basal ganglia have a key role in motivating procedural learning based on reward. Specifically, a cortico-striatal input is reinforced if it is associated with dopa-minergic input, which signals the upcoming reward57,60. The reinforced signal is used either to select the ongoing behavior or to be retained as a memory.
Our experimental data36 and model40 suggest that reinforcement might occur independently in separate loop
Quick and accurate performance might be achieved by the cerebellum
Skilful performance after long-term practice involves quick and coordinated movements of multiple joints and requires fine-tuning of movement parameters (such as velocity, force and timing). Our data suggest that such a learned motor skill depends on the loop circuit formed by the motor cortices and the anterior cerebellum (including the dorsal dentate nucleus41). This is consistent with a common view on the function of the cerebellum61.
According to our scheme (Box 1; Fig. 2), individual
Flexible ways to acquire new sequences
We have suggested that procedural learning proceeds as a gradual transition from a spatial sequence to a motor sequence. However, because these sequences are organized in a parallel fashion, either of these mechanisms could take the initiative to learn a sequential procedure. If explicit trial-and-error processes are not involved (as in a serial reaction-time task15), learning might be initiated by the motor sequence mechanism, rather implicitly69. The motor sequence mechanism could then guide
Concluding remarks
On the basis of behavioral and physiological experiments on trained monkeys and humans, we propose a scheme for the learning of sequential procedures. According to our scheme, the sensorimotor transformations required for individual actions (serial process) are replaced gradually with information on the sequence of actions (parallel process). Obviously, our scheme is incomplete and oversimplified, but it will certainly provoke further questions, such as those suggested below, which might be
Acknowledgements
The authors thank Johan Lauwereyns and Mitsuo Kawato for insightful discussion and comments. This study was supported by Grant-in-Aid for Scientific Research on Priority Areas from The Ministry of Education, Science and Culture of Japan, The Japan Society for the Promotion of Science (JSPS) Research for the Future Program and CREST (Core Research for Evolutional Science and Technology) of Japan Science and Technology Corporation (JST).
Selected references (74)
- et al.
Physiol Rev
(1974) - et al.
Biol Cybern
(1987) Nature
(1995)J Neurosci
(1998)Annu Rev Neurosci
(1997)- et al.
Curr Opin Neurobiol
(1998) Nat Neurosci
(1998)Exp Brain Res
(1997)- et al.
Annu Rev Neurosci
(1992) - et al.
J Neurophysiol
(1991)