| williams_et_al_cns_2019.pdf |
ABSTRACT:
Learning curves have long been used to explain changes in performance. Despite the depth of research investigating neural learning signals, whether changes in these signals follow learning curves is still untested. In the current research, we sought to determine the pattern of change that EEG-derived neural signals elicit across learning. In a reinforcement learning paradigm, participants learned sixty words of a novel language by developing symbol-word associations. Additionally, a reinforcement learning model conducted the same task. Trial-by-trial analyses for both empirical and simulated data were completed for accuracy rates, reaction times, and neural signals reflecting reward prediction errors (i.e., the reward positivity). We hypothesized that accuracy rates would increase and reaction times would decrease analogous to power-law learning curves. Further, we predicted that the reward positivity would too diminish in this fashion. Indeed, we found behavioural and neural measures of learning alike adhered to power law functions for both empirical and simulated data. Thus, for the first time we demonstrated that learning networks of the brain function similarly to well-known behavioural phenomena.
Learning curves have long been used to explain changes in performance. Despite the depth of research investigating neural learning signals, whether changes in these signals follow learning curves is still untested. In the current research, we sought to determine the pattern of change that EEG-derived neural signals elicit across learning. In a reinforcement learning paradigm, participants learned sixty words of a novel language by developing symbol-word associations. Additionally, a reinforcement learning model conducted the same task. Trial-by-trial analyses for both empirical and simulated data were completed for accuracy rates, reaction times, and neural signals reflecting reward prediction errors (i.e., the reward positivity). We hypothesized that accuracy rates would increase and reaction times would decrease analogous to power-law learning curves. Further, we predicted that the reward positivity would too diminish in this fashion. Indeed, we found behavioural and neural measures of learning alike adhered to power law functions for both empirical and simulated data. Thus, for the first time we demonstrated that learning networks of the brain function similarly to well-known behavioural phenomena.
SUPPLEMENTAL MATERIAL
Figure S1. Participant and model fits. Three trend fits for accuracy, reaction time, and reward positivity measures. Linear = ax+b, exponential = abx, power = axb. Goodness of fit measures reflect R-squared(GLMM) values.
Figure S2. Relationship between neural and behavioural measures for participant and all model simulated data. Top: relationship between reward positivity amplitudes and accuracy rates, bottom: relationship between reward positivity amplitudes and reaction times. Lines reflect grand averaged linear regression for participant and model data. Each participant and simulation contributed five data points for each plot – one for each of the first five trials. The intensity of colors scales to trial in that trial 1 is the darkest colours and trial 5 is the lightest colours.
Reinforcement Learning Model
Linear Mixed Effect Models
