Large-scale interactions in predictive processing: oscillatory versus transient dynamics

Abstract

How do the two main types of neural dynamics, aperiodic transients and oscillations, contribute to the interactions between feedforward (FF) and feedback (FB) pathways in sensory inference and predictive processing? We discuss three theoretical perspectives. First, we critically evaluate the theory that gamma and alpha/beta rhythms play a role in classic hierarchical predictive coding (HPC) by mediating FF and FB communication, respectively. Second, we outline an alternative functional model in which rapid sensory inference is mediated by aperiodic transients, whereas oscillations contribute to the stabilization of neural representations over time and plasticity processes. Third, we propose that the strong dependence of oscillations on predictability can be explained based on a biologically plausible alternative to classic HPC, namely dendritic HPC.

Keywords: alpha; beta; classical hierarchical predictive coding; dendritic hierarchical predictive coding; dynamics; efficient coding; feedback; feedforward; gamma; hierarchy; oscillations; plasticity; predictive coding; predictive processing; rhythms; synchronization.

Figure I. Relationship of layers to cortical rhythms.

(A) Laminar recordings from macaque V1 during attention task. A gamma source is present both in supra- and infragranular layers [113]. (B) Unit recordings of human alpha during eye closure shows strongest spike phase locking in L3 [118]. (C) Power spectra in [45]. Note that the reported gamma effect comprises broadband fluctuations. The reversed pattern of alpha/beta and high-frequency power is specific to unipolar recordings but not found in CSD and bipolar recordings, as illustrated by the bar plot on the right that visualizes the difference between unipolar and bipolar/CSDs [120, 16, 112].

Figure I. Consistency and interpretation of gamma feedforward, alpha/beta feedback patterns

(A) We argue that the positive correlation between gamma DAI (y-axis) and feedforward anatomical connectivity (x-axis) (observed by Bastos et al. [10]) is driven by V1 and V2, and is absent when V1 and V2 are not included. In the case of beta, we argue that the correlation is driven by DP and 7A. Significant Granger asymmetries are shown based on Figure S3 in Bastos et al. [10]. The areas are ordered hierarchically, with arrows from lower to higher nodes corresponding to feedforward connections. See also [120]. (B) We argue that strong Granger-causal influences at beta frequencies originate from regions with strong beta power, both in the feedforward [25, 26] and feedback direction[10]. (C)

Figure 1. Broadband fluctuations vs. narrow-band gamma.

In the main text, we have argued that narrow-band gamma oscillations are typically increased for predicted stimuli, while broadband ”gamma-frequency” activity increases for unpredicted stimuli. We argue that the broadband increase is explained by concurrent increases in spiking activity. In some systems like visual cortex, narrow-band gamma oscillations are frequently observed. Differences between Local Field Potential (LFP) spectra will reflect both the broadband and narrow-band gamma effect. By contrast, in other systems like auditory cortex, differences in LFP power predominantly reflect broadband fluctuations [53].

Figure 2. Emergence of rhythms in classic vs. dendritic hierarchical predictive coding.

In Classic HPC models, sensory inference results from interactions at each l-th hierarchical level between feedforward (FF) and feedback (FB) pathways that carry sensory prediction error (PE) and prediction signals, respectively [2, 1]. It was hypothesized that these feedforward error and feedback predictions signals are transmitted via 30-80Hz gamma oscillations in superficial layers, and alpha/beta oscillations in infragranular layers, respectively [8, 9, 10]. This hypothesis predicts strong gamma amplitude for unpredicted stimuli and weak amplitude for predicted stimuli, but does not entail a dependence of alpha/beta (see Main text). We argue that the emergence of gamma oscillations is better accounted for by the Dendritic HPC model, a biologically plausible PC model in which local E/I interactions play an important role [3]. The Dendritic HPC model builds on the anatomical observation that feedforward and feedback projections preferentially target basal and apical dendrites, respectively. Dendritic HPC does not contain specialized neurons (ϵ^l) for conveying feedforward error signals, and attributes error representation to voltage fluctuations in basal and apical dendrites. Error terms at the basal dendrites result from lateral inhibition that predicts and cancels out the feedforward inputs. Based on the Dendritic HPC model, we reason that a stimulus with high spatiotemporal predictability gives rise to tight E/I balance and sparse spiking activity, thereby promoting the emergence of fast network oscillations (as observed by [47]). We postulate that similar principles may also account for oscillations in other bands like beta and perhaps alpha.

Figure 3. Transients vs. oscillations in sensory inference.

We posit that sensory inference predominantly relies on aperiodic transients, while rhythms play a role in stabilizing neural representations and plasticity processes during the late, feedback-dominated phase of stimulus processing. In the visual system, a stimulus onset typically leads to rapid cascade of transients across the ventral stream. Stimulus onset latencies in IT are around 100ms and the inference of object properties has been largely completed around 120ms [72]. This is noted by the inference model q(z|x), the probability distribution over the latent variable z. Learning in supervised neural networks, but also in self-supervised neural networks like Variational Autoencoders, entails that feedback from higher levels reaches early sensory areas. Here the feedback may carry back-propagating errors that interact with local eligibility traces to instruct plasticity of the local recurrent connections [86]. Furthermore, for self-supervised learning, top-down generative networks can compute the posterior probability of the inputs given the inferred latents, p(x|z), which in turn instructs plasticity. Gamma oscillations in early visual areas emerge relatively late after stimulus onset, and will be prominent when this feedback arrives. As discussed in the text, gamma oscillations can facility plasticity processes, e.g. via synchronizing neurons that receive spatiotemporally correlated inputs [47], or via activation-dependent gamma-phase shifting [87]. We also argue that gamma may stabilize the neural activity during the later stimulus phases, narrowing the region of state space that is occupied.