Top-Down Inference in the Auditory System: Potential Roles for Corticofugal Projections

Abstract

It has become widely accepted that humans use contextual information to infer the meaning of ambiguous acoustic signals. In speech, for example, high-level semantic, syntactic, or lexical information shape our understanding of a phoneme buried in noise. Most current theories to explain this phenomenon rely on hierarchical predictive coding models involving a set of Bayesian priors emanating from high-level brain regions (e.g., prefrontal cortex) that are used to influence processing at lower-levels of the cortical sensory hierarchy (e.g., auditory cortex). As such, virtually all proposed models to explain top-down facilitation are focused on intracortical connections, and consequently, subcortical nuclei have scarcely been discussed in this context. However, subcortical auditory nuclei receive massive, heterogeneous, and cascading descending projections at every level of the sensory hierarchy, and activation of these systems has been shown to improve speech recognition. It is not yet clear whether or how top-down modulation to resolve ambiguous sounds calls upon these corticofugal projections. Here, we review the literature on top-down modulation in the auditory system, primarily focused on humans and cortical imaging/recording methods, and attempt to relate these findings to a growing animal literature, which has primarily been focused on corticofugal projections. We argue that corticofugal pathways contain the requisite circuitry to implement predictive coding mechanisms to facilitate perception of complex sounds and that top-down modulation at early (i.e., subcortical) stages of processing complement modulation at later (i.e., cortical) stages of processing. Finally, we suggest experimental approaches for future studies on this topic.

Keywords: auditory; colliculus; cortex; descending; medial geniculate body; speech perception; thalamus; top-down.

Figure 1

(A,B) Illustration of the experiment by Perrot et al. (2006) showing electrical stimulation sites in the human auditory cortex in panel (A). Black circles = auditory cortex stimulation sites, gray circles = non-auditory cortex stimulation sites, Roman numerals correspond to the individual patients. CA-CP, plane passing through the anterior and posterior commissures; VCA, vertical plane passing through the anterior commissure; VCP, vertical plane passing through the posterior commissure. Panel (B) shows the change in the variation in the amplitude of evoked otoacoustic emissions (EOAEs) under spontaneous conditions (dark bar), after stimulation in the non-auditory cortex (gray bar), and after stimulation in the auditory cortex (white bar). These data illustrate that human auditory cortical stimulation diminishes the variability of evoked otoacoustic emissions. **P < 0.01; ***P = 0.001; NS, not significant using paired t-tests. Standard error of the mean is shown using error bars. Data obtained with permission from Perrot et al. (2006). Panel (C) Illustrates the impact of attending to an auditory stimulus on distortion product otoacoustic emissions (DPOAEs). As shown, attending to an acoustic stimulus diminishes the DPOAE amplitude (red trace), compared to ignoring that stimulus (green trace). Data obtained with permission from Smith et al. (2012). (D) Illustration of a model proposed by Terreros and Délano (2015)to explain the influence of the cortex on the cochlea. They propose multiple potential pathways from the auditory cortex, involving the inferior colliculus and the superior olivary nuclei, to impact the outer hair cells via the medial olivocochlear bundle. Figure obtained with permission from Terreros and Délano (2015).

Figure 2

Illustration of the principles of Bayesian inference in neuronal coding. Top-down pre-conceived notions about a sound (illustrated as a sound trace surrounded by a blue haze) are combined with noisy information from the periphery (illustrated as a noisy sound trace entering the ear). Bayes theorem (in the box) combines the pre-conceived notions with the noisy sensory information to recover the original signal (here represented as the posterior probability or the actual percept.

Figure 3

Generic example of the simplest circuit to involve top-down modulation to implement Bayesian predictive coding. A top-down projection (in green) carries the predictive signal [P(x)] from Figure 2. Such a signal could be derived, for example, from the frontal cortex or auditory cortex. This descending input is combined with weighted information from the periphery (represented as I) at an intermediate structure, such as the auditory cortex or medial geniculate body, using the examples provided above. Using this same scheme, I would be derived from the medial geniculate body or inferior colliculus. The weighting is determined by the reliability of the signal, conceived as a presynaptic input onto the input terminals. Neurophysiologically, this reliability signal could be represented by cholinergic or monoaminergic inputs that scale with arousal or attention. Note that this generic model is not limited to the structures listed on the figure, which are given as examples.

Figure 4

Diagram of known corticofugal and other subcortical descending projections across sensory systems. Black arrows, bottom-up projections; Blue arrows; top-down projections; CN, cochlear nuclei; DCN, dorsal column nuclei; IC, inferior colliculus; LGN, lateral genicular nucleus; MGB, medial geniculate body; NLL, nuclei of the lateral lemniscus; NTS, nucleus tractus solitarius; PAG, periaqueductal gray; PBN, parabrachial nuclei; SC, superior colliculus; SO, superior olive; VPL, ventral posterior lateral nucleus of the thalamus; VPM, ventral posterior medial nucleus of the thalamus. Taken with permission from Lesicko and Llano (2017).

Figure 5

Putative circuit motifs that can implement top-down modulation of frequency receptive fields using descending projections. Green tuning curves represent the modified tuning curves after descending projections were activated. Left, the simplest “gain control” motif whereby top-down projections dial the responsiveness of target cells up or down, thus shifting the frequency tuning curve up or down. Middle, a lateral inhibition motif, whereby descending input inhibits inputs representing frequencies other than the characteristic frequency. In this case, the tuning curve would sharpen. Right, an input selection motif, where top-down inputs would either enhance (denoted with a “+”) or suppress (denoted with a “−“) certain classes of inputs either pre-or post-synaptically. In doing so, the top-down projections could eliminate inputs from what was the previous characteristic frequency and thus shift the tuning curve laterally. Bottom corresponds to descending inputs synchronously eliciting excitatory post-synaptic potentials (EPSPs) in two neurons of different characteristic frequencies. If a sound object has multiple frequency components peaking at different times (e.g., an early low-frequency peak and a later high-frequency peak, marked with blue and red circles on the spectrogram, respectively), then when those bottom-up inputs arrive at neurons with synchronized EPSPs, they are more likely to fire synchronous spikes, thus linking them as part of one auditory object. CF, characteristic frequency.