Feedforward and feedback interactions between visual cortical areas use different population activity patterns

Abstract

Brain function relies on the coordination of activity across multiple, recurrently connected brain areas. For instance, sensory information encoded in early sensory areas is relayed to, and further processed by, higher cortical areas and then fed back. However, the way in which feedforward and feedback signaling interact with one another is incompletely understood. Here we investigate this question by leveraging simultaneous neuronal population recordings in early and midlevel visual areas (V1-V2 and V1-V4). Using a dimensionality reduction approach, we find that population interactions are feedforward-dominated shortly after stimulus onset and feedback-dominated during spontaneous activity. The population activity patterns most correlated across areas were distinct during feedforward- and feedback-dominated periods. These results suggest that feedforward and feedback signaling rely on separate "channels", which allows feedback signals to not directly affect activity that is fed forward.

Fig. 1. Studying feedforward and feedback interactions using neuronal population activity.

a Each circle represents a neuron in each area, with the shading representing the activity level of the neuron. The population activity patterns involved in feedforward signaling (top) might be distinct from those involved in feedback interactions (bottom). b Schematic showing a sagittal section of occipital cortex and the recording setup for the V1–V2 recordings. We simultaneously recorded V1 population activity using a 96-channel array and V2 population activity using a set of movable electrodes and tetrodes. c Schematic showing an overhead view of the recording setup for the V1–V4 awake recordings. We simultaneously recorded V1 and V4 population activity using one 96-channel and one 48-channel array in V1 and a 48-channel array in V4 in the first animal, and two 96-channel arrays in V1 and two 48-channel array in V4 in the second animal.

Fig. 2. Using Canonical Correlation Analysis (CCA) to capture population interactions.

a Relating activity across two neuronal populations. Each circle represents the population activity recorded on a given trial. For each activity point observed in the V1 population (left panel; gray dots), there is a corresponding, simultaneously recorded activity point observed in V2 (right panel, gray dots). The red axes represent the first pair of canonical dimensions, identified using CCA. Neuronal activity projected onto the first pair of canonical dimensions (red dots) is highly correlated across the two areas (bottom panel). b Spike counts across the recorded neurons are taken in specified time windows (gray boxes), which may either be positioned at the same time in both areas (i.e., t₁ = t₂) or with a delay between areas (t₁ ≠ t₂). The activity in each gray box is represented by a circle in panel (a). c The population correlation function corresponds to the correlation between areas returned by CCA (the correlation associated with the first pair of canonical dimensions), as a function of the time delay between areas (t₂ − t₁).

Fig. 3. V1–V2 interaction transitions from feedforward-dominated shortly after stimulus onset to feedback-dominated during the spontaneous period.

a Inter-areal zero-delay population correlation increased throughout the trial, and was higher for spontaneous activity than for evoked activity. Zero-delay refers to spike counts taken in the same time window in the two areas (t₁ = t₂ in Fig. 2b). Black line shows the average across all recording sessions for which the V1 and V2 populations have aligned receptive fields. Shading indicates S.E.M. Dotted line shows average across all recording sessions where the V1 and V2 receptive fields are misaligned. Gray line shows average population correlation after shuffling trial correspondence between the two areas. b Population correlation functions for an example session (red: early evoked, yellow: late evoked; purple: spontaneous). Faded lines show population correlation functions after shuffling trial correspondence between the two areas (note that there are multiple superimposed lines). c Population correlations at all times during the trial. The horizontal axis represents the time delay between areas (t₁ − t₂), and the vertical axis represents time relative to stimulus onset (t₁). Horizontal lines (red, yellow, and purple) indicate epochs used in b. Dashed vertical line indicates zero-delay population correlations shown in a. White area denotes times for which population correlations could not be computed: the V2 activity window had reached either the beginning or the end of the trial. Same session as in b. d Feedforward ratio for different epochs of evoked and spontaneous activity. Left panel shows sessions for which the V1 and V2 populations have aligned receptive fields; right panel shows sessions where the V1 and V2 receptive fields are misaligned. Feedforward ratio is defined as the difference between the area under the feedforward (positive delay) and feedback (negative delay) sides of the population correlation function, divided by their sum. Solid symbols show the average across all recording sessions, whereas open symbols correspond to each recording session. Insets show average feedforward ratios after shuffling trial correspondence between the two areas for each recording session (horizontal lines show 1 S.D. intervals, most of which are not visible because they are smaller than the width of the plotted symbol). e An early feedforward peak is only present in recording sessions where the V1 and V2 populations have aligned receptive fields. Peak height is measured after performing a jitter-correction to isolate fast timescale interactions (see Methods and Supplementary Fig. 3). Circles correspond to recording sessions for which the V1 and V2 populations have aligned receptive fields. Triangles correspond to sessions in which the V1 and V2 receptive fields are misaligned.

Fig. 4. V1–V4 interaction transitions from feedforward- to feedback-dominated during the evoked period.

a Inter-areal zero-delay population correlation increased throughout the evoked period. Black line shows average across all recording sessions. Shading indicates S.E.M. Gray line shows average population correlations after shuffling trial correspondence between the two areas. b Population correlation functions for an example session, for early (red) and late evoked (yellow) activity. Due to the short duration of the inter-stimulus period, we could not compute a population correlation function for spontaneous activity. Faded lines show population correlation functions after shuffling trial correspondence between the two areas (note that there are multiple superimposed lines). c Population correlations at all times during the trial. The horizontal axis represents the time delay between areas (t₁ − t₂), and the vertical axis represents time relative to stimulus onset (t₁). Horizontal lines (red and yellow) indicate epochs used in b. The dashed vertical line indicates zero-delay population correlations shown in a. White area denotes times for which population correlations could not be computed: the V4 activity window had reached either the beginning or the end of the trial. Same session as in b. d Feedforward ratio for early and late evoked activity. Solid circles show the average across all recording sessions, whereas open circles correspond to each recording session. Insets show average feedforward ratios after shuffling trial correspondence between the two areas for each recording session (horizontal lines show 1 S.D. intervals, most of which are not visible because they are smaller than the width of the plotted symbol).

Fig. 5. Illustration of how to assess whether feedforward- and feedback-dominated interactions involve the same population activity patterns.

a Canonical dimensions identified during a feedforward-dominated period in the trial (red dimensions). These are putative “Feedforward” (FF) canonical dimensions. Open red circles denote activity during the feedforward-dominated period. Solid red circles denote the projection onto the FF canonical dimensions. b We can then ask whether these FF canonical dimensions generalize to a feedback-dominated period. One possibility is that the interaction structure (defined using the canonical dimensions) remains stable across the two periods. In this case, the FF canonical dimensions (red dimensions) capture a similar level of correlation during the feedback-dominated period as the canonical dimensions identified during this period, the putative “Feedback” (FB) canonical dimensions (blue dimensions). As a result, the normalized correlation, the ratio of the population correlation for the FF canonical dimensions to that for the FB canonical dimensions (both computed in a cross-validated manner; see Methods), is close to 1. Open blue circles denote activity during the feedback-dominated period. Solid purple circles denote the projection of activity during the feedback-dominated period onto the FF canonical dimensions. Solid blue circles denote the projection onto the FB canonical dimensions. c Alternatively, the interaction structure might change across the two periods. In this case, the FF dimensions capture only a small fraction of the population correlation during the feedback-dominated period. Same conventions as in b.

Fig. 6. Interaction structure is distinct for the feedforward- and feedback-dominated periods.

a The dimensions found by fitting CCA shortly after stimulus onset (80 ms after stimulus onset) do not generalize well to later epochs in the evoked period, and worse still during the spontaneous period. Gray lines correspond to each of the 5 recording sessions. We report the normalized correlation, defined as the total correlation captured at the test epoch by the dimensions fit to some other epoch over the total correlation captured by the dimensions fit the test epoch (both computed in a cross-validated manner; see Methods). b Dimensions identified late in the evoked period (1180 ms after stimulus onset) do not generalize well to early evoked epochs and to epochs in the spontaneous period, but generalize well to mid-evoked activity. Same conventions as in a. c Dimensions identified during the spontaneous period do not generalize well to the evoked period. Same conventions as in a. d Assessing changes in interaction structure across the entire trial. The trial was divided into 100 ms segments, and CCA was applied separately to the activity in each time window. The top two canonical pairs associated with each window were then used to capture inter-areal correlations in the other time windows (see Methods). Each row corresponds to the time during the trial during which the canonical dimensions were identified. Each column corresponds to the time during the trial where the population correlation is assessed. Each entry shows the average across all recording sessions. Straight arrow highlights the comparison of the interaction structure within the evoked period. Curved arrow highlights the comparison of the interactions structure between the spontaneous and the evoked periods. Dashed white boxes indicate epochs reproduced in f. e Comparing identified dimensions across epochs for the awake V1–V4 recordings. The trial was divided into 100 ms segments, and CCA was applied separately to the activity in each time window. The top canonical pair associated with each window was then used to capture inter-areal correlations in the other time windows (see Methods). Arrow highlights the comparison of the interaction structure within the evoked period. Same conventions as in d. f Detailed view of the V1–V2 generalization performance for the comparable epochs between the V1-V2 and V1-V4 recordings. Epochs are indicated by the dashed white boxes in d.

Fig. 7. Summary of results.

During the early evoked period, interactions between areas tend to be feedforward-dominated. Later during the evoked period and during the spontaneous period, interactions between areas become feedback-dominated. Furthermore, feedforward- and feedback-dominated interactions involve different population activity patterns. Larger ellipses represent the set of all activity patterns one might observe in either the V1 or the V2/V4 populations. The smaller ellipses represent the activity patterns most related across the two areas.