Experience-Dependent Adaptation of the Spatial Generalization of Human Motor Adaptation KA Thoroughman and JA Taylor, Dept. of Biomedical Eng., Washington University, St. Louis, MO Humans possess a remarkable ability to learn new motor skills, such as adapting to the dynamics of a novel object. Motor learning of novel dynamics has been characterized by experiments in which human participants learn, over many movements, to counter forces generated by a robotic manipulandum1. Recent experiments have characterized, for the first time, a more fundamental adaptation: the transformation of sensation during single movements into changes in predictive control in immediately subsequent movements2. These movement-by-movement learning steps have proven to be surprisingly broad in their generalization across position3 and velocity4,5 space, suggesting that the tuning of kinematic parameters underlying motor adaptation is also broad. The novel dynamics presented to humans, however, have been simple enough to be learned with broad movement-by-movement generalization. Here we train participants to move in three environments with differing complexity to determine whether the breadth of generalization can itself adapt in response to the environment. Twelve subjects performed reaching tasks toward 16 visually displayed targets while holding a manipulandum. On day one, the subjects did not experience any forces. On each of days two, three, and four, subjects were exposed to a single viscous force field whose force magnitude was proportional to hand speed. The force direction depended on hand velocity direction. On different days, we altered the spatial complexity of the forces such that the force direction could change as fast (Field One), twice as fast (Field Two), or four times as fast (Field Four) as the velocity direction (Figure 1). Each force field training day had a total of 640 movements. In the first half of the training set, the force field was always present, but during the latter half, subjects experienced occasional catch trials in which the field was unexpectedly removed. During the first half, we quantified learning by calculating the correlation coefficient of the velocity time series to each participant’s day one movements. We found that subjects improve performance in all fields. In comparing Field One to Field Two we find that the improvement (p>0.9) and asymptotic level (p>0.1) of performance were similar, but that the improvement (p<0.04) and asymptotic level (p<0.001) of performance in Field Four is significantly different than the other fields. The similarity in learning seen in the first half of training belied a striking difference in how subjects learn movement-by-movement in Fields One and Two. We used state-space analysis to identify the trial-by-trial evolution of the participants’ internal dynamic model5:
y n = DFn − z n
z n +1 = z n + b( DFn − z n ) The input (F) is a two-dimensional vector of the direction of the force field. The hidden state (z) is updated by the approximation of the actual position error (DF-z) multiplied by a sensitivity parameter (b). The dependence of the sensitivity parameter on the angular difference between input and output movement directions identifies the spatial extent of generalization in movement-by-movement learning (Figure 3). The generalization function in Field Four is relatively small, flat, and unremarkable. The generalization function of Field One is very broad and always positive, such that an error sensed in one direction always generates the same sign (positive or negative) of adaptation in all subsequent movement directions. The generalization function of Field Two, however, is significantly narrower (p<0.02) and, unlike in Field One (p<0.001), features negative components in movement directions far away from sensed error. This “Mexican-hat” feature of generalization is interestingly inappropriate for Field Two, as movements separated by 180 degrees have equal force directions; we speculate that this sign flipping of adaptation may reflect a stereotypical way the brain can encode narrow tuning of movement kinematics. We therefore have identified, for the first time, how mere minutes of training can inform the meta-adaptation of human transformation of sense into action: the spatial grain of movement-bymovement learning narrowed in response to a narrower spatial grain of environmental dynamics. This remarkable change calls into question models of motor learning such as radial basis function networks6 in which the encoding of movement kinematics is typically held fixed to facilitate rapid and linear solutions to nonlinear problems inherent to learning dynamics.
KA Thoroughman and JA Taylor m=1
m=2
0
-0.4
0.4
y velocity (m/s)
0.4
y velocity (m/s)
y velocity (m/s)
0.4
m=4
0
-0.4
-0.4
0 x velocity (m/s)
0.4
0
-0.4 -0.4
0 x velocity (m/s)
0.4
-0.4
0 x velocity (m/s)
0.4
Figure 1: Force fields for the spatial frequencies examined in the study. The forces are always proportional to the magnitude of the subject hand Cartesian velocity, but the direction of the force varies with respect to the spatial frequency m. Left: Field One (in which m=1); center: Field Two (m=2); right: Field Four (m=4). In each field, Fx = 15rv sin(mθ v ) and Fy = −15rv cos(mθ v ) , where rv and θv are the magnitude and direction of hand velocity. Correlation between null field and force field hand velocities 0.96 0.94
Figure 2: Correlation coefficients comparing hand velocities in the force fields to baseline (day one, no force) performance. Movements made in a particular direction in the force field were compared with the baseline trajectory in that particular direction. Data are averaged across subjects (n=12) and was smoothed using a rectangular window across 20 movements.
Correlation Coefficient r
0.92 0.9 0.88 0.86 0.84 Field One Field Two Field Four
0.82 0.8
0
50
100
150 200 Movement Number
250
300
350
Figure 3: Sensitivity parameter (b) for each force field across target directions. The x-axis represents the angular separation (degrees) between input and output movements and the y-axis is a measure of the sensitivity of adaptation to previously sensed error. Field One data are in blue, Field Two data are in green, and Field Four data are in red. The data are averaged across all subjects (n=12).
Generalization Parameter
0.25
Field One Field Two Field Four
0.2
0.15
b
0.1
References: 1. Shadmehr R, Mussa-Ivaldi. FA. J Neurosci, 14:3208 (1994). 2. Scheidt RA et al. J Neurophys, 86:971 (2001). 3. Hwang EJ et al. PLoS Biol, 1:E25. (2003). 4. Thoroughman KA, Shadmehr R. Nature, 407:742 (2000). 5. Donchin O et al. J Neurosci, 23:9032 (2003). 6. Pouget A, Snyder LH. Nature Neurosci, 3:1192 (2000).
0.05
0
-0.05
-0.1 -200
-150
-100
-50 0 50 Angular Separation (deg)
100
150
200