Marisa Birk¹ (), Annika Tesio¹, Ingmar de Vries^1,2, Moritz Wurm¹; ¹CIMeC - Center for Mind/Brain Sciences, University of Trento, Italy, ²Donders Institute, Radboud University, Nijmegen, The Netherlands

According to the theory of predictive coding, the brain constantly generates predictions about sensory input. These predictions are compared against the incoming signal, with accurate predictions suppressing expected bottom-up information. In this framework, stimuli that are more predictable lead to stronger predictive representations and more suppressed sensory input. By contrast, unpredictable stimuli result in weaker top-down predictions and stronger, less suppressed representation of the bottom-up input. Here, we tested these hypotheses using dynamic Representational Similarity Analysis (dRSA). This approach uses temporally variable models that capture the representational content of a dynamic event at each individual time point. This allows testing when the brain represents a stimulus at a given time point in relation to its real-time occurrence, i.e., in a predictive or lagged manner. This provides a means to disentangle top-down predictions and bottom-up sensory input. During magnetoencephalography (MEG), 22 participants watched videos of a moving dot. We manipulated the predictability of the dot’s trajectory by varying the probability that the dot would change its direction, creating four conditions that ranged from fully predictable to highly unpredictable. dRSA was applied to capture neural representations of the dot's position and movement direction. For all predictability levels, we found that dot position was represented with a consistent lag of 120 ms. By contrast, the dot’s direction was represented in both predictive (-1100 to -300 ms) and lagged (900 to 1700 ms) manners. Importantly, the strength of dRSA for both predictive and lagged representations systematically increased with higher predictability. Thus, while top-down predictions became more pronounced with increasing predictability, we did not observe a suppression of lagged (potentially bottom-up) representations, challenging a key assumption of predictive coding. Instead, our findings suggest that increased predictability sharpens the representation of bottom-up sensory input by enhancing the selectivity for relevant stimulus features.