Neural basis of sensorimotor learning: modifying internal models

The neural basis of the internal models used in sensorimotor transformations is beginning to be uncovered. Sensorimotor learning involves the modification of such models. Different stages of sensory-motor processing have been explored with a continuum of experimental tasks, from learning arbitrary associations of sensory cues to movements, to adapting to altered kinematic and dynamic environments. Several groups have been studying changes in neuronal activity in cortical and subcortical areas that may be related to the acquisition and consolidation processes. We discuss the progress and challenges in understanding how these learning-related neural changes are involved in the modification of internal models, and offer future directions.

Introduction

Sensorimotor learning is essential for daily behavior. It is used for maintaining known behavioral abilities and for learning new skills. For example, it is used to learn how to move a cursor on a computer screen or to assign behavioral relevance to some arbitrary stimulus, such as the instruction ‘stop the car’ when the traffic light is red. Current opinion holds that sensorimotor actions are produced by transformations that utilize internal models of the mechanics of the body and the world [1] (Figure 1b). Inverse models work in the obvious direction: transforming desired goals into a plan to accomplish them^a [2] (Figure 1a). These transformations are undoubtedly a major part of sensorimotor control. However, since action based solely on sensory feedback is slow and risky, the brain — like efficient artificial control systems — has developed the power of prediction. Predictive abilities can be achieved by using forward models that simulate the outcomes of a given plan (Figure 1a). There are compelling theoretical arguments why forward models are an essential feature of sensorimotor control, perception, and learning [3]. First, forward models can be used to give instantaneous predicted feedback that provides an estimate of the state of the controlled effector, and thus overcome the significant delays of real sensory feedback. Therefore, forward models are a necessary ingredient for the contemporary description of the sensorimotor system as an optimal feedback controller [4]. When real sensory feedback does arrive, it can be combined with such predicted feedback to compensate for limitations and noise of the sensory systems [5]. A second proposed role for forward models involves anticipating and cancelling the sensory effects of self initiated actions from the incoming sensory stimuli (reafference). Finally, during learning itself, forward models may be used to generate sensory error signals (predicted feedback minus real feedback) which can guide learning of inverse models (distal teacher) [6]. An increasing number of psychophysical experiments support the notion that humans make use of forward models (e.g. [7, 8, 9, 10], for a review see [11]). Classic experiments on fish [12, 13], bats [14], and cats [15] have demonstrated the existence of corollary discharges (the outputs of the forward models) and their effects on reafference. More recently, two elegant series of electrophysiological experiments in monkeys have culminated in showing how forward models are used in the saccadic eye movement system [16^••], and in sensing of head motion in the vestibular system [17]. Mulliken et al. [18^•] showed that during hand movements a subpopulation of neurons in the posterior parietal cortex of monkeys may encode a forward estimate of the cursor direction on the screen. They found neurons whose maximal encoding evolved too late to subserve feedforward motor processing, but too early to result from sensory feedback.

The extent that prediction is used by different brain functions remains an open question. The theory of active perception proposes that in order to perceive the world, predictions can direct active exploratory movements (e.g. of the eyes, or the vibrissa [19]) and top-down attention to the informative features of the stimulus. Going even further, the proposal that perception itself is none other than the process of integrating our internally generated predictions with the incoming sensory stream is an attractive one. In this vain, Noë argues that ‘the experience of seeing occurs when the organism masters what we call the governing laws of sensorimotor contingency’ [20]. Thus, forward models may be a more fundamental and pervasive feature of brain organization than just a component of motor control [21] (Figure 1c).

Different sensorimotor learning tasks modify the relevant internal models. Depending on the context, a red light may be associated with different behaviors. Likewise, we may need to learn different sensorimotor transformations when we control a cursor by using different devices, such as a mouse, a touch-pad, or an electronic-pen. When the design of such devices merges well with our internal models, we learn to control novel devices more quickly and naturally.