1.5 The organisation of the dissertation
1.5.2 Interaction between the learning processes
In the second part of the thesis we investigated how motor memories of a dynamic task formed in association with different contexts interacted with one another. Previous studies have extensively examined the context dependent properties of motor memories, illustrating
1.5 The organisation of the dissertation 27
how contexts (e.g., movement direction, workspace location, background colour, etc) affect the allocation of motor memories to different tasks (e.g., Howard et al., 2012, 2013). However, very few have examined the interaction of motor memories associated with different contexts.
Theoretical studies of motor learning that aim to describe multi-context adaptation mostly assume a fixed interaction between context-dependent adaptive processes (Donchin et al., 2003; Ingram et al., 2013; Kim et al., 2015b; Lee and Schweighofer, 2009). This is normally done by assuming a fixed tuning function of adaptive memories across different contexts. As such, when a given context is presented, the adaptive memory associated with that context is updated the most, and other adaptive memories are updated to a lesser extent depending on the distance between their corresponding context and the presented context.
In the current study, we look more closely into the interaction of context-dependent motor memories. In chapter 5, we show that when a novel dynamic task is learned under two distinct contexts, the interaction between the associated motor memories can increase or decrease depending on the environmental condition. This behaviour is accounted for by a novel state-space model that features an adaptive coupling factor to capture the dynamics of the interaction between two motor memories. In chapter 6, we examine the properties of such interaction when subjects learn a task with a familiar dynamics (as opposed to a novel dynamics) under two different contexts. We illustrate that in the case of familiar dynamics, the learned motor memories exhibit a consistent interaction.
Part II
SENSORIMOTOR LEARNING IN 3D
ENVIRONMENTS
Chapter 2
Spatial generalisation of adaptation in 3D
2.1 Introduction
A fundamental question in sensorimotor control is how the brain generalises an acquired skill or experience from one context to another. This is particularly important for the brain to be able to perform as well in similar situations without having to learn specific sensorimotor commands for every single state of the body or environment. Generalisation has been studied in various ways in sensorimotor control, from perception to action (for example, Liu and Weinshall, 1999; Roach et al., 2017; Sarwary et al., 2015; Shadmehr, 2004). In particular, there are many studies investigating the extent to which learning a new skill or a new sensorimotor mapping in one context transfers to others. In this case, various experimental paradigms, such as force field learning, visuomotor adaptation, or object manipulation are used to examine the generalisation behaviour in various contexts (Donchin et al., 2003; Ingram et al., 2010; Wu and Smith, 2013). A common aspect of generalisation in such paradigms has been the tuning of a learned dynamics to movement direction in a 2D environment. For example, many studies have looked into how learning a force field in a particular movement direction generalises to other directions (Berniker et al., 2014;
Donchin et al., 2003; Mattar and Ostry, 2007; Mussa-Ivaldi et al., 1994; Shadmehr, 2004;
32 Spatial generalisation of adaptation in 3D
Fig. 2.1 A. The pattern of generalisation to different target directions (data from Howard and Franklin (2015)). Adaptation is maximum at the training target (φ = 0◦) and decays for targets farther away. B. Schematic illustration of the transfer of force field learning from one movement direction (Y axis of the main frame) to another (y axis of the target-based frame).
The learned adaptive force in the training direction (green arrow) can be transferred to the new orientation (φ ) in two possible frame representations: the target-based frame (purple arrow) or the extrinsic frame (blue arrow). The force measurement by channel trials is limited to the lateral dimension of the movement direction (that is, the x axis of the target-based frame). Thus, if learning is transferred in the extrinsic frame (blue arrow) only its lateral component can be measured by channel trials.
Thoroughman and Shadmehr, 2000). It has been shown that subjects exhibit the maximum adaptation in the training direction, and their adaptive response decays, typically in a Gaussian manner, as the movement direction deviates from the training direction (Fig. 2.1A; Brayanov et al., 2012; Howard and Franklin, 2015). This typical decaying generalisation is also observed in other contexts such as different movement speeds or target distances (Goodbody and Wolpert, 1998; Mattar and Ostry, 2010).
Most studies on generalisation have been conducted by the use of planar robots, which helped to obtain the tuning curves of adaptation transfer over different movement directions, or arm configurations on the horizontal plane (Berniker et al., 2014; Howard and Franklin, 2015). The interpretation of such tuning patterns, however, has not been easy as two major factors were involved in obtaining these patterns. First, the internal representation of the coordinate systems in which the generalisation occurs is still an open question (Berniker et al., 2014; McDougle et al., 2017). For instance, when force field adaptation is transferred
2.1 Introduction 33
from one direction to another, it could be transferred in an extrinsic world based frame, or in an intrinsic target-based frame, or a mixture of two (Fig. 2.1B). Each of these possibilities suggests a different way of interpretation of the obtained tuning curves. Second, the mea-surement of adaptation in planar robots is limited to a single dimension (that is, the lateral dimension of the movement direction), which may not match the actual direction in which generalisation happens (the direction of generalisation is not clear due to unknown coordinate frame of representation). For example, in Fig. 2.1B, two different coordinate frames are compared for the transfer of learning from one target direction to another. As shown, for the target-based frame (x-y frame), the direction of generalisation (purple arrow) is orthogonal to the direction of movement, and thus matches the direction of measurement by channel trials. For the extrinsic frame (XY frame), however, the direction of transfer (blue arrow) is aligned with the direction of compensation in the original target direction (the green arrow).
In this case, only the lateral component of the transferred force can be measured (that is, the component that is perpendicular to the movement direction), and thus the generalisation cannot be fully examined.
In this study, we used the 3BOT to address this issue by examining the target-based versus extrinsic (world-based) transfer of learning. Subjects adapted to a velocity dependent force field while performing reaching movements towards a training target at the intersection between the sagittal and transverse planes (Fig. 2.2A). We probed the transfer of learning when subjects reached to several targets distributed in the 3D space with different reach angles compared to the training target (Fig. 2.2A). The 3D distribution of probe targets was designed such that for one group of targets, both extrinsic and target-based representations predicted the same pattern of generalisation. Whereas, for another group of targets, different representations predicted different generalisation patterns. Comparing the measured transfer of learning between these groups determined the coordinate frame in which the generalisation had occurred.
34 Spatial generalisation of adaptation in 3D