Shaping the Cortical Landscape: Functions and Mechanisms of Top-Down Cortical Feedback Pathways

Abstract

Cortical feedback pathways are proposed to guide cognition and behavior according to context and goal-direction. At the cellular level, cortical feedback pathways target multiple excitatory and inhibitory populations. However, we currently lack frameworks that link how the cellular mechanisms of cortical feedback pathways underlie their cognitive/behavioral functions. To establish this link, we expand on the framework of signal routing, the ability of cortical feedback pathways to proactively modulate how feedforward signals are propagated throughout the cortex. We propose that cortical feedback modulates routing through multiple mechanisms: preparing intended motor representations, setting the trigger conditions for evoking cortical outputs, altering coupling strengths between cortical regions, and suppressing expected sensory representations. In developing this framework, we first define the anatomy of cortical feedback pathways and identify recent advances in studying their functions at high specificity and resolution. Second, we review the diverse functions of cortical feedback pathways throughout the cortical hierarchy and evaluate these functions from the framework of signal routing. Third, we review the conserved cellular targets and circuit impacts of cortical feedback. Fourth, we introduce the concept of the "cortical landscape," a graphical depiction of the routes through cortex that are favored at a specific moment in time. We propose that the cortical landscape, analogous to energy landscapes in physics and chemistry, can capture important features of signal routing including coupling strength, trigger conditions, and preparatory states. By resolving the cortical landscape, we may be able to quantify how the cellular processes of cortical feedback ultimately shape cognition and behavior.

Keywords: attention; cortical circuits; feedback; functional connectivity; motor preparation; predictive coding.

Figure 1

Early descriptions of feedback pathways in subcortical and cortical structures. These drawings of neural circuits by Santiago Ramón y Cajal are among the first depictions of feedback pathways. (Left) Ramón y Cajal’s drawing of retinal circuitry. Arrows in the downward direction (arrows “g” and “e”) reflect feedforward or centripetal pathways from the photoreceptors towards higher-order brain structures for the transmission of visual stimuli. Ramón y Cajal additionally noted the presence of feedback or centrifugal fibers (arrow “a”) which originate in higher-order brain areas and target the retina. Thus, feedback occurs even at the earliest levels of sensory processing. (Right) Ramón y Cajal’s drawing of pathways of the somatosensory system. Demarcated are feedforward fibers from the brainstem to thalamus (arrow “G”) and from the thalamus to the cortex (arrow “b”). Also illustrated here are feedback fibers from cortex to thalamus (arrow “a”). While feedback fibers appear to be relatively sparse near the periphery, they are highly abundant in the thalamocortical system. In describing this anatomical organization during his Nobel prize speech in 1906, Ramón y Cajal speculated as to the function of feedback: “What is the role of the centrifugal fibres? Are they, as Duval thinks, conductors destined to produce in the sensory pathway articulations a very intimate contact which would be indispensable for the satisfactory propagation of the nervous impulse? Or rather do they transport some form of energy from the brain, the rapid accumulation of which in the sensory stations is necessary for the passage of ascending nerve currents? Unfortunately, at this stage of science, it is impossible to give satisfactory and categorical answers to these questions”. Now, over 100 years later, scientists continue to debate the functions of feedback fibers (panels and quote are reproduced with permission from © The Nobel Foundation: https://www.nobelprize.org/prizes/medicine/1906/cajal/lecture/).

Figure 2

Diverse functions of cortical feedback throughout the hierarchy. (Left) Hierarchical organization of the primate visual system adapted with permission from Felleman and Van Essen (1991). Circled regions are some of the cortical regions discussed in the manuscript. (Right) Proposed functions of cortical feedback, segregated according to source and the target region. (Red) prefrontal cortex (PFC): Brodmann area 46, dorsolateral PFC; (green) premotor cortices: FEF, frontal eye field; DPc, dorsal premotor cortex; MPc, medial premotor cortex; M2, secondary motor cortex; (yellow) higher-order sensory cortices: IT, infratemporal cortex; V4, visual area V4; MT, middle temporal visual area; (blue) lower-order sensory cortices: V1, primary visual cortex; S1, primary somatosensory cortex; A1, primary auditory cortex.

Figure 3

Feedforward and feedback functions of motor cortex. Motor cortex (center) projects both to subcortical structures along corticofugal pathways (solid arrow) and to sensory cortices along cortical feedback pathways (dashed arrow). These two pathways originate from different populations of motor cortex pyramidal neurons. However, within local regions of the motor cortex, there is alignment between the target movement field of the corticofugal neurons and the target receptive field of the cortical feedback neurons. Functions of the corticofugal pathway relate to motor control, including movement initiation, set-point modulation, and amplitude modulation. Functions of the cortical feedback pathway relate to spatial attention and corollary discharge.

Figure 4

Models of predictive coding. (A) A framework of predictive coding based on movement initiation. A motor command signal (bottom left) triggers motor execution (solid arrow) and sends a copy of the motor command to sensory cortices (efference copy or corollary discharge, dashed arrow). The forward model converts the motor command to a pattern of predicted sensory input. This predicted sensory input is subtracted from the feedforward sensory input, canceling out the reafferent (self-generated) component. Unexpected signals, not predicted by self-generated movement, are not subtracted and therefore propagate forward. Efference and prediction signals are carried by feedback pathways (dashes arrows) whereas sensory signals are carried by feedforward pathways (solid arrows). (B) A framework of predictive coding based on sensory expectation. Sensory expectation evokes signals reflecting the pattern of predicted sensory input, to subtract from afferent sensory input. Unexpected signals, not predicted by sensory expectation, are not subtracted and therefore propagate forward. Prediction signals are carried by feedback pathways (dashes arrows) whereas sensory signals are carried by feedforward pathways (solid arrows).

Figure 5

Cellular and circuit mechanisms of cortical feedback modulation. Proposed cellular targets and mechanisms of cortical feedback pathways. Blue axons and terminals represent cortical feedback pathways, yellow axons and terminals represent feedforward pathways. Neurons in the target area include pyramidal neurons (black) and multiple types of GABAergic interneurons (green and red). (A) Cortical feedback forms excitatory synapses onto pyramidal neurons, leading to increased excitation. (B) Cortical feedback synapses onto interneuron-targeting GABAergic interneurons (red), leading to dis-inhibition. (C) Coincident inputs from feedforward and feedback pathways lead to dendritic spiking and supra-linear increases in pyramidal neuron output. (D) Phasic activity of cortical feedback leads to pyramidal neuron synchronization, both locally (shown here) and between source and target regions (not shown). Synchronization can ensure optimal post-synaptic integration and thereby enhance coupling among synchronized regions. (E) Cortical feedback synapses onto pyramidal neuron-targeting GABAergic interneurons, leading to increased inhibition and pyramidal neuron decorrelation. It is likely that any feedback pathway may deploy any combination of these mechanisms, as needed for a specific goal or context.

Figure 6

Energy landscapes across disciplines. (A) Gibbs free energy curve illustrating the transformation of reactants to products across an energetically unfavorable transition state. This curve depicts a reaction occurring within a single dimension. (B) Reactions occurring within two dimensions can be depicted as an energy landscape or potential energy surface. Shown here are three simulations (arrows) along a hypothetical two-dimensional energy landscape. The initial conditions and contours of the landscape determine the simulation trajectories, which flow from states of high (peaks) to low (troughs) free energy. (C) A simulated energy landscape for the re-folding of denatured green fluorescent protein (GFP), with multiple potential pathways (arrows) undertaking different partial folding states. (D–F) An analogous form of energy landscapes, used here to model neural circuit activity. (D) The structure of the circuit model is two populations (S1 and S2) which self-excite (arrows) and cross-inhibit (circles). Each population can be driven by external inputs (I1 and I2). (E) A phase plane representation of the activities of S1 vs. S2 provides an activity landscape of the network. Critical features of this landscape are the fixed points demarcated at the crossings of the nullclines (orange and green lines). Stable fixed points (black dots) are analogous to troughs in an energy landscape. The unstable fixed point in this example (gray dot) would have the shape of a saddle in an energy landscape: sloping down towards the fixed point (black arrows towards) and sloping down away from the fixed point (black arrows away). The two simulations shown in blue and red have the same starting point but diverge from each other and converge into the two stable fixed points. (F) Changing the inputs alters the landscape, which now (I1 > I2) favors activity trajectories towards the top-left fixed point. Panel (C) is reproduced with permission from Reddy et al. (2012), panels (E,F) are reproduced with permission from Wong and Wang (2006), copyright 2006 Society for Neuroscience.

Figure 7

Feedback shapes the cortical landscape according to context. Framework for how feedback pathways may contribute to flexible sensory-motor processing. Top panels illustrate hierarchically-arranged cortical nodes with input regions at the bottom and output regions at the top. Blue arrows represent sensory inputs (bottom) or cortical outputs (top). Blue nodes represent highly active regions whereas yellow nodes represent less active regions. Black lines represent the structural and functional connectivity between nodes, with thicker lines depicting stronger connectivity. Bottom panels are energy landscape representations of the above networks. The height of each contour reflects the energy barrier of propagating into that region. (A) If signal processing between cortical regions is defined only by feedforward connections, then sensory-motor processing will always follow the strongest connections. (B) Different contexts, such as goal-direction or self-generated movement, can propagate along feedback and laterals pathways [green lines (excitatory) and red lines (inhibitory)]. Consequently, these top-down signals modulate which input and output nodes are enhanced or suppressed, thereby altering the coupling strengths between cortical regions (black lines). The same input signal may now be routed to diverse outputs, as appropriate for the current context. As illustrated in the bottom panels, modulations of activity levels within individual regions and coupling strengths between regions change the contours of the cortical landscape, favoring different input-output trajectories.