Behavioural Study of the Effect of Trial and Error Versus Supervised Learning of Visuo-Motor Skills

Introduction

Acquiring a sequential skill typically requires compiling a number of elementary actions into a unique chain that forms the complete sequence. Human skill learning has been extensively studied Thut et al., 1996; Perez et al., 2007; Rosenbaum, Kenny and Derr, 1983; Grafton et al., 1998; Jueptner et al., 1997a, 1997b; Sakai et al., 1998; Hikosaka et al., 1996, 2000; Bapi et al., 2000. Way back in 1964, Fitts observed that subjects are more attentive in the initial phase of learning a skill and become more automatic in the later stages of learning when attention can be engaged in other tasks, such as performance of dual tasks. Various aspects of skill learning have been studied in monkeys using a 2 × 5 sequence learning paradigm (summarized in Hikosaka et al. 1995, 1996) in which a sequence of 10 button presses is learned by trial and error. These experiments showed that as training progressed, monkeys improved on two measures of performance: errors—monkeys made fewer errors before attaining a success criterion—and response time (RT), the time taken to perform a sequence decreased with training. While errors reached a minimum level within a shorter period of training, RTs continued to improve over longer periods (Hikosaka et al. 1995, 1996). Bapi et al. (2000, 2006) followed up this in humans studying whether the skills are tied to the effectors or they are abstract. They concluded that in the initial stage of learning, effector-independent representation in visual/spatial coordinates is formed and this transforms to an effectordependent representation in motor coordinates by the late stage and there are distinct brain areas sub-serving these two representations.

Another issue of importance in skill learning is the learning strategy or paradigm employed. Learning paradigms can be grouped broadly into two main categories: supervised and unsupervised, based on whether evaluative feedback was provided or not. Feedback signal provides an assessment of the performance of the system during the learning process. In supervised learning, we assume that the teacher provides the desired response at each instant of time that can be used to calculate the errors and make appropriate corrections in order to eventually achieve the desired target. In a variation of supervised learning called reinforcement learning, a coarse feedback indicating the quality of the output is provided without specifying the desired response itself. In unsupervised learning, the desired response is not known. Thus explicit error information cannot be used to improve behaviour in unsupervised learning. The system needs to discover the inherent regularities present in the inputs and self-organize the information. Thus adaptive sequential skills could be acquired using any one of the strategies, that is, supervised, reinforcement, or unsupervised learning. The existing studies were not explicitly designed to tease out differences between trial and error and guided modes of learning motor sequences.

In the current study we investigated the effects of two learning paradigms: supervised (explicit guidance) and reinforcement (trial and error). We adopted a modified version of 2 × 5 visuo-motor sequence learning task (Hikosaka et al., 1995; Bapi et al., 2000) wherein subjects learned the same number of finger movements in two learning modes. Supervised learning mode involves learning the sequential dependency structure by following a series of visual cues. On the other hand, trial and error learning requires an active exploration of visual cues, working memory for previous choice of responses, evaluation of responses and learning the sequence structure (Figure 4.1). In this chapter, we emphasize issues related to analysis of learning related effects in the behavioural data. The aim is to recommend a methodology for characterizing a subject's stage and extent of learning of a skill. We have also acquired functional MR images while subjects engaged in the sequence learning tasks. The imaging results will be discussed elsewhere.

...