Learning to Learn: Environmental Consistency Modulates Motor Adaptation Rates L. Nicolas Gonzalez Castro*1,2, Matthew Hemphill*1, Maurice Smith1. The human motor system has the ability to adapt to different physical environments. This motor adaptation acts to reduce future motor errors, however this ability depends on the consistency of the environment. Here we explore whether human subjects can not only adapt their motor output but also adapt the rate at which this adaptation occurs. This would allow environment‐specific adaptation that could improve motor performance beyond the level possible with a static learning rate. In a highly inconsistent environment, the motor system would benefit from a low learning rate in order to avoid responding to current disturbances that do not predict future disturbances. However, in a highly consistent environment, the motor system would benefit from learning at a higher rate because disturbances experienced on one trial should be highly predictive of future disturbances – and thus any learning associated with the current disturbance is likely to improve future performance. We tested this prediction by exposing subjects to different learning environments and measuring the learning rates associated with each environment.
In these environments the consistency of force‐field application was systematically manipulated by varying the duration of repeated blocks during which the force‐field was activated (and occasionally inverted). The environments are diagrammed in Figure 1. In the first environment (P/N1; 3 subjects) subjects were exposed to a single positive force‐field trial followed by a negative force‐field trial, followed by 7‐9 washout (null field) trials. In the second environment (P1; 12 subjects) subjects were exposed to a single positive force‐field trial followed by 7‐9 washout (null field) trials. And in the third and fourth environments (P7 and P20; 12 subjects each), subjects were exposed to 7 and 20 force‐field trials, respectively, followed by 14‐16 and 27‐29 washout trials, respectively.
The single‐trial learning rates associated with these environments were measured by inserting error‐clamp (EC) trials before and after the first force‐field (FF) trial in a random subset (44%) of the learning blocks. These measurement triplets (EC‐FF‐EC) are diagrammed as dashed black vertical lines Figure 1. The difference between the lateral force profiles during the post‐FF and pre‐FF error‐clamp trials was used to compute the learning rate. Hand paths as well as mean angular errors assessed at the peak speed point are shown for the initial force‐field trials in each environment in Figure 2. The errors induced on these trials did not vary significantly between environments (one‐way ANOVA, F(3,17) = 1.32, p = 0.28).
Despite the similarity between the motor errors caused by these force‐field trials, the learning rates associated with these trials were significantly different between environments (one‐way ANOVA, F(3,17) = 7.1 , p<0.003 ) and varied by more than a factor of ten. Figure 3 shows the change in lateral force between the post‐FF and pre‐FF produced by subjects in the different learning environments and Figure 4 shows the associated learning rates (i.e., the fraction of the force‐field compensated on the post‐FF trial). These learning‐induced changes in lateral force after single‐ trial FF exposure reached levels as high as 0.40 of the force‐field magnitude during the latter part of the P20 environment which had an average learning rate of 0.28. In contrast, learning rates averaged 0.17 in the P7 environment, 0.062 in the P1 environment, and 0.026 in the P/N1 environment.
Although the learning rates observed after the first encounter with the learning stimuli early in the experiments are not significantly different (one‐way ANOVA, F(3,17) = 2.0, p = 0.15), the last learning rate measured in each environment shows clear statistical differences (one‐way ANOVA, F(3,17) = 16.7, p < 0.0001). As shown in Figure 5 learning rates in the two more consistent environments (P7 & P20) appeared to increase with subsequent exposures to the stimuli. This indicates that the increases in learning rates associated with these environments progress slowly and gradually. These results show that learning rates are readily modulated by the statistics of the environment and that these modulations appear to reflect both increases and decreases from baseline adaptation rates. Further work is needed to delineate the mechanisms that control these modulations. * These authors contributed equally to this work. 1. Harvard School of Engineering and Applied Sciences, Cambridge, MA 2. Harvard‐MIT Division of Health Sciences and Technology, Cambridge, MA
1
2 Mean Trajectory During First Trial of Learning Stimulus
Learning Stimuli 1
P/N1 P1 P7 P20 P/N1 Peak Speed P1 Peak Speed P7 Peak Speed P20 Peak Speed
0
0
0
-0.01
P/N1
Testing
-1
20
40
60
80
100
120
140 -0.02
1 0
-0.03
20
40
60
80
100
120
140
1 0 Testing
-1 0
P7 20
40
60
80
100
120
140
-0.04
Mean Angular Error @ Peak Speed 20
-0.05
18 16
Angular Error (deg)
0
Vertical Displacement (m)
P1
Testing
-1
-0.06
-0.07
1 0 Testing
-1 0
-0.08
P20 20
40
60
80
100
120
14 12 10 8 6 4 2
140
-0.09
Trial Number
P/N1
P1
P7
P20
Learning Stimulus -0.1 -0.02
-0.01
0
0.01
0.02
0.03
Horizontal Displacement (m)
3
4
5
Mean Learning Rates for the Different Learning Stimuli
Learning‐induced Changes in Force Profiles
Evolution of Learning Rates 0.6
1.8
P/N1 P1 P7 P20
1.6 1.4
0.4
Late Exposure
0.4
1 0.8 0.6
Learning Rate
0.3
Learning Rate
Force (N)
1.2
P/N1 P1 P7 P20
0.5
All Trials
0.2
0.3 0.2 0.1
0.1 0.4 0
0.2
0 -0.1
0 -0.2
-0.1 0
200
400
600
Time (ms)
800
1000
1200
P/N1
P1
Learning Stimulus
P7
P20
-0.2
0
2
4
6
8
10
12
Measurement
14
16
18
20