Mapping Shape to Visuomotor Mapping: Learning and Generalisation of Sensorimotor Behaviour Based on Contextual Information

Abstract

Humans can learn and store multiple visuomotor mappings (dual-adaptation) when feedback for each is provided alternately. Moreover, learned context cues associated with each mapping can be used to switch between the stored mappings. However, little is known about the associative learning between cue and required visuomotor mapping, and how learning generalises to novel but similar conditions. To investigate these questions, participants performed a rapid target-pointing task while we manipulated the offset between visual feedback and movement end-points. The visual feedback was presented with horizontal offsets of different amounts, dependent on the targets shape. Participants thus needed to use different visuomotor mappings between target location and required motor response depending on the target shape in order to “hit” it. The target shapes were taken from a continuous set of shapes, morphed between spiky and circular shapes. After training we tested participants performance, without feedback, on different target shapes that had not been learned previously. We compared two hypotheses. First, we hypothesised that participants could (explicitly) extract the linear relationship between target shape and visuomotor mapping and generalise accordingly. Second, using previous findings of visuomotor learning, we developed a (implicit) Bayesian learning model that predicts generalisation that is more consistent with categorisation (i.e. use one mapping or the other). The experimental results show that, although learning the associations requires explicit awareness of the cues’ role, participants apply the mapping corresponding to the trained shape that is most similar to the current one, consistent with the Bayesian learning model. Furthermore, the Bayesian learning model predicts that learning should slow down with increased numbers of training pairs, which was confirmed by the present results. In short, we found a good correspondence between the Bayesian learning model and the empirical results indicating that this model poses a possible mechanism for simultaneously learning multiple visuomotor mappings.

Relative errors can cue absolute visuomotor mappings

“…For learning associations between colour cues and rotated visuomotor mappings, there is evidence that learning is dependent on whether the participant had explicit knowledge about this relationship (Hegele and Heuer 2010). Also for learning visuomotor shifts, there is evidence that awareness of the cues plays a dominant role for whether or not learning occurs (van Dam and Ernst 2015). The same could be true when there are no associative cues, as in the present study.…”

“…That is, there were no additional associative cues indicating the required mappings in an a priori fashion as was the case in many previous studies (e.g. Kravitz 1972;Kravitz andYaffe 1972, 1974;Martin et al 1996;Seidler et al 2001;Hay and Pick 1966;Pick et al 1969;Donderi et al 1985;van Dam et al 2013;van Dam and Ernst 2015;Welch 1971). Without associative contexts, there is in principle no need for any long-term learning to occur, since the feedback is only provided after the movement is already complete and thus cannot provide a priori knowledge about the currently correct mapping.…”

“…Results were consistent with Mapping Learning despite the relative nature of the error information in the feedback. This shows that (Donderi et al 1985;Fernandez-Ruiz et al 2000Kravitz 1972;Kravitz andYaffe 1972, 1974;Martin et al 1996;Pick et al 1969;Seidler et al 2001;van Dam et al 2013;van Dam and Ernst 2015;Welch 1971). Such associative cues help us to identify the correct mapping that should be used even before a movement is initiated.…”

Optimal visual–haptic integration with articulated tools

“…In principle, previously learned tool mappings could be identified via a range of cues, all of which were available in our experiments. For example, ‘contextual’ cues, such as the colour/appearance of the tool, could allow switching before movement even begins (Imamizu et al 2007; van Dam and Ernst 2015a). Similarly, static spatial information about hand opening and the tool tips could allow identification.…”

Value generalization in human avoidance learning

“…Stimuli were 5-fold radially symmetric flower-like shapes, as described in van Dam and Ernst, 2015. These were selected on the basis of previous psychophysical testing demonstrating that they are perceptually linear, and the fact that they can be individually generated along axis a single perceptual axis of ‘spikiness’ using the mathematical description provided in the paper.…”