Generating stimulus anticipation from stimulus and prediction history

International audience ; Perceptual decisions can be defined as a motor action performed subsequent to sensory stimulation. Alongside the type of stimulus, prior expectations about the upcoming stimulus and past performance impact behavior in terms of confidence [1], response times [2] and accuracy [3]. In decision models, such as the drift-diffusion model, pre-stimulus states are represented as the starting point of evidence-accumulation. However, since the mechanisms of prediction are mainly unconscious [4], little information about this starting point can be extracted from data at the single-trial level. Therefore, a generative model of anticipation could enhance our understanding of single decisions. Here we assume that the pre-stimulus state depends on past predictive states as well as on stimuli, and propose five Bayesian networks updated sequentially, that either rely on past stimuli or past predictions and stimuli. Previous studies have shown that decisions in fact depended at least marginally on past stimuli and decisions [5]. These effects have already been modelled by means of a modification of psychometric curves [6] or Bayesian inference [7], but only considered past stimuli. Since the observation of prediction in empirical data is not straightforward, we also developed four metrics to assess the quality of each model, based on response times and accuracy of responses. In addition, we compared our results to the starting point distribution predicted by the drift-diffusion model [8].We show that models taking into account past accuracy of prediction in addition to past stimuli match data better in terms of response times and accuracy than models that do not. Our results moreover suggest that prediction variability relates to the variance of starting point distributions fitted by the drift-diffusion model.Our results provide a new basis for understanding the origin of prediction and suggest that prediction performance should also be considered when inferring the next stimulus. Future work could ...