Neural signatures of predictive coding are context-dependent
Speaker: Dr. Ryszard Auksztulewicz (Department of Biomedical Sciences, City University of Hong Kong)
Date: 2017.12.1 (Fri) 15:00-17:00
Location: 20F Conference room , DA-AN Branch, Taipei Medical University

The brain is thought to generate internal predictions to optimise behaviour. However, it is unclear to what extent these predictions are modulated by other top-down factors such as attention and task demands. In this talk I will present results of three studies combining human electrophysiology and computational modelling to identify the neural mechanisms of sensory predictions and their interactions with current context.

First, using magnetoencephalography (MEG) and dynamic causal modelling (DCM), sensory predictions and temporal attention were orthogonally manipulated in an auditory mismatch paradigm, revealing interactive effects on evoked response amplitude. This interaction effect was modelled in a canonical microcircuit using DCM. While mismatch responses were explained by recursive interplay of sensory predictions and prediction errors, their attentional modulation was linked to increased early sensory gain.

Second, we analysed electrocorticographic data recorded from patients performing a task orthogonally manipulating “what” and “when” predictability of auditory targets. The two predictability types modulated evoked responses in different cortical regions and at dissociable latencies. DCM served to disambiguate between models of stimulus predictability in terms of top-down processing and gain modulation: “what” predictability increased auditory short-term plasticity, while “when” predictability increased putative synaptic gain in motor areas. This suggests that distinct predictions are mediated by qualitatively different neural mechanisms.

Finally, we independently manipulated the spatial/temporal predictability of visual targets, and the relevance of spatial/temporal information provided by auditory cues. Relevance modulated the influence of predictability on task performance. To explain these effects, we estimated our participants’ subjective predictions using a Hierarchical Gaussian Filter (HGF). Model-based time-series of predictions and prediction errors were linked to dissociable induced activity measured with MEG. Predictions correlated with beta-band activity, while prediction errors were signaled by increased gamma and decreased alpha-band activity. Crucially, these oscillatory correlates were modulated by task relevance, suggesting that current goals influence prediction signaling.