Kanai 2015

Cerebral hierarchies: predictive processing, precision and the pulvinar

This paper considers neuronal architectures from a computational perspective and asks what aspects of neuroanatomy and neurophysiology can be disclosed by the nature of neuronal computations? In particular, we extend current formulations of the brain as an organ of inference—based upon hierarchical predictive coding — and consider how these inferences are orchestrated. In other words, what would the brain require to dynamically coordinate and contextua- lize its message passing to optimize its computational goals? The answer that emerges rests on the delicate (modulatory) gain control of neuronal populations that select and coordinate ( prediction error) signals that ascend cortical hierarchies. This is important because it speaks to a hierarchical anatomy of extrinsic (between region) connections that form two distinct classes, namely a class of driving (first-order) connections that are concerned with encoding the content of neuronal representations and a class of modulatory (second-order) connec- tions that establish context—in the form of the salience or precision ascribed to content. We explore the implications of this distinction from a formal perspec- tive (using simulations of feature–ground segregation) and consider the neurobiological substrates of the ensuing precision-engineered dynamics, with a special focus on the pulvinar and attention.

--

These asymmetries include the laminar specificity of forward and backward connections, the prevalence of nonlinear or modulatory backward connections (that embody interactions and nonlinearities inherent in the generation of sensory signals) and their spectral characteristics — with fast (e.g. gamma) activity predominating in forward connections ( prediction errors) and slower (e.g. beta) frequencies emerging as this evidence is accumulated in units that provide descending predictions [20 – 22].

Computationally, this gain corresponds to the precision or confidence associated with ascending prediction errors;

In this setting, hidden causes in the generative model control the precision or variance of subordinate causes generating data.

Crucially, prediction errors are affected by descending predictions in one of two ways: expectations can either generate predictions of first-order effects, through the functions g(μ). Alternatively, they can generate predictions of precision, through the functions π(μ)

Key Sentence #1, from caption of Figure 3.
Crucially, these constraints are modulated by top-down predictions of their precision (blue arrows).

The above mentioned examples illustrate a form of precision or gain control that is intrinsic to the cortical hierarchy and speaks to separate descending streams of prediction—that predict the first and second-order attributes of lower-level representations.

In contrast, the predicted precision has an excitatory modulatory effect,

These findings support the notion that neurons in the pulvinar encode expected precision or confidence in information used for perceptual decisions.

The pulvinar’s contribution to gain control has been demonstrated in a compelling study of spike-field coherence [57]. By concurrently recording pulvinar spikes and local field potentials from V4 and TEO, the authors showed that the spike-field coherence between the pulvinar neurons and alpha oscillation in V4 and TEO was enhanced when attention was directed to the receptive field of the pulvinar neuron. Crucially, conditional Granger causality analysis across the three regions showed that the pulvinar neurons facilitated the trans- mission of information between V4 and TE by synchronizing the alpha oscillation in those cortical regions.

These studies provide neurophysiological evidence that the pulvinar neurons encode expected precision, and modulate the gain of corticocortical communication. The notion of precision engineering in the pulvinar offers a coherent (computational) perspective on how seemingly disparate aspects of attention (gain modulation) and confidence (uncertainty estimation) are orchestrated. Although the concepts of salience, confidence and attention may appear distinct, their intimate relationship can be interpreted as an integral part of perceptual inference—reflecting the different faces of precision.