Correlations of neural predictability and information transfer in cortex and their relation to predictive coding
Predictive-coding like theories agree in describing top-down communication through the cortical hierarchy as a transmission of predictions generated by internal models of the inputs. With respect to the bottom-up connections, however, these theories differ in the neural processing strategies suggested for updating the internal model. Some theories suggest a coding strategy where unpredictable inputs, i.e., those not captured by the internal model, are passed on through the cortical hierarchy, whereas others claim that the predictable part of the inputs is passed on. Here, we addressed which neural coding strategy is employed in cortico-cortical connections using an information-theoretic approach. Our framework allows for quantifying two core aspects of both strategies, namely, predictability of inputs and information transfer, through local active information storage and local transfer entropy, respectively. A previous study on the neural processing of retinal ganglion cells connected to the lateral geniculate nucleus showed a coding for predictable information, captured by an increase in the information transfer with the predictability of inputs. Here, we further investigate predictive coding strategies at the cortical level. In particular, we analyzed LFP activity obtained from intracranial EEG recordings in humans and spike recordings from mouse cortex. We detected cortico-cortical connections with increasing information transfer with the predictability of inputs in recorded channels from frontal, parietal and temporal areas in human cortex. In the mouse visual system, we detected connections exhibiting both an increase and decrease in the information transfer with input predictability, although the former was predominant. Our evidence supports the presence of both predictive coding strategies at the cortical level, with a potential predominance of encoding for predictable information. SummaryThe ability of the brain to infer the hidden causes of sensory experiences has been conceptualized within the computational framework of predictive coding. This framework explains perceptual inference and learning as a process of constantly updating an internal model of the world. Predictive coding describes cortical activity as a communication of sensory evidence and predictions generated from prior expectations. While different views of predictive coding agree on the communication of prior expectations throughout cortex, they differ in how internal expectations are updated. One view states that, to update the internal model, the cortex propagates the mismatch between the expected neural activity and the actual neural response to sensory stimuli. In contrast, another view suggests that the cortex propagates the match between the expected neural activity and the actual neural response. In this work, we were able to tease apart these two views, both in human cortex and the mouse visual system, using information theory. We observed that the brain predominantly propagates expected information, i.e., the match between prior expectations and incoming sensory inputs.