Non-refereed supplementary results for
Human EEG uncovers latent generalizable rule structure during learning
Collins AGE, Cavanagh JF, Frank MJ
a. behavioral results
Asymptotic learning phase errors. We report here the supplementary behavioral results that replicate previous findings and confirm model predictions, further supporting the fact that subjects in this experiment build task-set structure. Errors during the asymptotic learning phase were indicative of task-set structure. In particular, errors were significantly more likely following switches of the high dimension (t=6.13, p<10e-4), than following switches of the low dimension (t=2.34, p=0.025). This error switch cost is consistent with task-switching effects, and further validates the use of asymmetric RT switch cost for identifying the higher order dimension. Moreover, after high dimension switches, errors of neglect of the high dimension (NH) were significantly faster than errors of neglect of the low dimension (NL) (first iteration of the task, p=0.0024, all three blocks p=0.0038 figure 2d), consistent with our models in which NH errors can arise from impulsive re-application of the previous task-set (Collins & Frank, 2013). Most importantly, the effect of switch vs. stay H, but not L, interacted with error type. There were significantly more NL errors than NH errors for switch H (t=4.52, p=0.0001), but not for stay H (p=0.82, difference t=4.8, p<10e-4), suggesting that most task-set switch errors corresponded to a successful task-set retrieval, but unsuccessful task-set use for the current stimulus. Conversely, for the low dimension, NL errors were overall more numerous than NH errors, irrespective of switch or stay L (both p's<0.01, difference p=0.74), showing successful maintenance of the task-set, but erroneous use of it. This is the same pattern we obtained in our previous experiment(Collins & Frank, 2013), and confirms asymmetrical roles of the two input dimensions, with H controlling a hierarchically higher level of task-set.
b. EEG results
Supplementary ERP components. P2 component was defined as averaged voltage over peak at [224-236] ms, N2a and N2b components at [280-320] ms and [340-352] ms respectively, and P3 at [404-416] ms. MPFC components were quantified as N2b-P2 and N2a-P3, but similar results are obtained with N2a or N2b.
Figure 1: Cue-locked ERP effects. Top half: Structure specific switch effects over right dorso-lateral prefrontal cortex. Top row: Late negativity. a) ERP over right dorso-lateral prefrontal (rdlPFC) electrodes for switches and stays on the High and Low dimensions. b) average value over late window 480-580ms shows main effect of Switch-H vs. Stay-H (p=0.008), but no effect of Switch-L vs. Stay-L (p>0.7). c,d) t-values for contrasts of switching on high and low dimensions (p<0.01 uncorrected), revealing that these late ERP effects occur over mid- and right anterior sites, including the rdlPFC ROI (ellipse overlay on c). Bottom row: Early positivity effects. e) The amplitude difference of Switch-H vs. Stay-H, a marker of hierarchical task structure during learning, is correlated with subsequent positive transfer. f,g) Topographic maps of significant (p<0.01) correlations between ERP indices and positive transfer reveal that these correlations are selective to Switch-H, compared to both Stay-H and to Switch-L, and are localized to right anterior electrodes and left posterior sites (reversed coefficients potentially suggest dipolar effects).
Bottom half: Structure independent switch effects over medial prefrontal cortex. a) ERP over medial prefrontal (mPFC) electrodes for switches and stays on the High and Low dimensions. Top row: N2P2 component. b) N2P2 component shows main effect of Switch vs. Stay for both high and low dimensions. c,d) t-values for contrasts of switching on high and low dimensions (p<0.01 uncorrected), revealing that these late ERP effects occur over mid-anterior sites, including the mPFC ROI (ellipse overlay on c). Bottom row: N2P3 effects. E) N2P3 component shows main effect of Switch vs. Stay for both high and low dimensions, with a significant interaction. e,f) t-values for contrasts of switching on high and low dimensions (p<0.01 uncorrected), revealing that these late ERP effects occur over mid-anterior sites, including the mPFC ROI.
mPFC represents non task-set specific conflict and change. The N2 component (figure 1, third row) showed main effects of switch vs. stay for both the high and the low dimension, centered over medial prefrontal cortex. We thus focused on the a priori defined compound mPFC electrode. For N2 (quantified as P2N2 peak to trough measure, see methods), we obtained significant main effects of switch vs. stay for both H and L (p<10e-3, t>3.7), but no interaction (p>0.36). In particular, SwitchH-StayL trials were not different from StayH-SwitchL (t = 0.3, p=0.74), showing that a change in a single input dimension was treated similarly irrespective of the dimension. However, all other pairwise comparisons were significant (t>2.1, p<0.04). In particular, a change in both input dimensions was significantly different from a change in only one of H or L dimensions. These findings suggest independent and additive effects of switch in both dimensions, indexing an amount of change in the input.
The P3 component (see figure 1, last row), also showed significant main effects of Switch vs. stay for both H and L again (t>2.7, p<0.01), however this time the two factors interacted strongly (t = 3.4, p=0.002). This interaction was qualified by a difference between repeat trials from all three other conditions (t=3.8, p<0.0005), but no difference between any of those three non-repeat conditions (t<1, p>0.3). This is in agreement with the usual interpretation of a P3 component as indexing surprise (Mars et al., 2008). Indeed, stay trials show the strongest P3 component, with no difference for other conditions, as they are three times less likely than non-stay trials.
Collins, A. G. E., & Frank, M. J. (2013). Cognitive control over learning: Creating, clustering, and generalizing task-set structure. Psychological Review, 120(1), 190–229. doi:10.1037/a0030852
Mars, R. B., Debener, S., Gladwin, T. E., Harrison, L. M., Haggard, P., Rothwell, J. C., & Bestmann, S. (2008). Trial-by-trial fluctuations in the event-related electroencephalogram reflect dynamic changes in the degree of surprise. The Journal of neuroscience : the official journal of the Society for Neuroscience, 28(47), 12539–45. doi:10.1523/JNEUROSCI.2925-08.2008