Brief Stimuli Cast a Persistent Long-Term Trace in Visual Cortex

Abstract

Visual processing is strongly influenced by recent stimulus history, a phenomenon termed adaptation. Prominent theories cast adaptation as a consequence of optimized encoding of visual information by exploiting the temporal statistics of the world. However, this would require the visual system to track the history of individual briefly experienced events, within a stream of visual input, to build up statistical representations over longer timescales. Here, using an openly available dataset from the Allen Brain Observatory, we show that neurons in the early visual cortex of the mouse indeed maintain long-term traces of individual past stimuli that persist despite the presentation of several intervening stimuli, leading to long-term and stimulus-specific adaptation over dozens of seconds. Long-term adaptation was selectively expressed in cortical, but not in thalamic, neurons, which only showed short-term adaptation. Early visual cortex thus maintains concurrent stimulus-specific memory traces of past input, enabling the visual system to build up a statistical representation of the world to optimize the encoding of new information in a changing environment.SIGNIFICANCE STATEMENT In the natural world, previous sensory input is predictive of current input over multisecond timescales. The visual system could exploit these predictabilities by adapting current visual processing to the long-term history of visual input. However, it is unclear whether the visual system can track the history of individual briefly experienced images, within a stream of input, to build up statistical representations over such long timescales. Here, we show that neurons in early visual cortex of the mouse brain exhibit remarkably long-term adaptation to brief stimuli, persisting over dozens of seconds, and despite the presentation of several intervening stimuli. The visual cortex thus maintains long-term traces of individual briefly experienced past images, enabling the formation of statistical representations over extended timescales.

Keywords: long-term adaptation; mouse; sensory adaptation; sensory encoding; visual cortex; visual processing.

Figure 1.

Visual cortex and thalamus exhibit orientation-specific adaptation to the immediately preceding (1-back) grating. A, Schematic of Neuropixels probe insertion trajectories through visual cortical and thalamic areas, adapted with permission from Siegle et al. (2021). B, Presentation sequence of drifting grating stimuli. Mice were shown drifting gratings with a duration of 2 s, separated by a 1 s gray screen. Gratings were drifting in one of eight different dire ctions (0, 45, 90, 135, 180, 225, 270, 315°) and were presented in random order. For the analysis of orientation-specific adaptation, we contrasted activity to gratings preceded by gratings of the same orientation (repeat, blue) with that elicited by gratings preceded by a grating of the orthogonal orientation (orthogonal, red). C, Population peristimulus time histograms of neurons in V1 for repeat and orthogonal conditions. The transient response is reduced when the same orientation is successively repeated, indicating orientation-specific adaptation. Subsequent analyses focused on this transient response (0–100 ms, green shaded area). Vertical dashed lines denote stimulus onset and offset, respectively. Bin width = 25 ms. Error bars indicate SEM. D, One-back adaptation ratios of transient responses across visual areas. Adaptation ratios were computed by dividing the firing rate of each neuron for repeat by that for orthogonal stimulus presentations and therefore express the response magnitude to a repeated stimulus orientation relative to that elicited by the same stimulus orientation but preceded by a grating with the orthogonal orientation. Adaptation ratios smaller than one indicate adaptation. All visual areas show significant 1-back adaptation. Error bars indicate bootstrapped 95% confidence intervals. White numbers indicate the number of neurons in each area. E, The average firing rate to a stimulus preceded by a stimulus with the same orientation (x-axis) is consistently smaller than the firing rate to a stimulus preceded by a stimulus with the orthogonal orientation (y-axis) across mice (gray dots denote different mice; size scaled by the number of neurons of each mouse) in both thalamus (left) and cortex (right), as indicated by data points positioned above the diagonal. F, Histograms of single-neuron adaptation ratios (log transformed) in thalamus (left) and cortex (right). Negative x values indicate adaptation, and the red dashed line marks zero adaptation (i.e., equal firing rates for repeat and orthogonal conditions). The triangle shape indicates the mean adaptation across the population of neurons with p-value indicating the significance of the population mean. AL, Anterolateral; AM anteromedial, LM, lateromedial; PM, posteromedial; RL, rostrolateral.

Figure 2.

Adaptation depends on orientation tuning and adaptor/test orientation. A–C, Orientation tuning curves in V1 for units of low (A), medium (B), or high (C) orientation selectivity (tertile split; see above, Materials and Methods), following adaptation to different 1-back grating orientations (colored arrows). Stimulus and adaptor orientations are expressed relative to the preferred orientation of each neuron. Tuning curves show local response reductions to the adapted orientation. D–F, Adaptation ratios as a function of the adaptor and test orientation relative to the neuron's preferred orientation. For instance, the adaptation ratio for a relative stimulus orientation of 0° compares the visual response to a test grating with the preferred orientation of the neuron when it is preceded by an adaptor grating with the same (preferred) orientation, versus when it is preceded by the orthogonal (nonpreferred) adaptor orientation (illustration in A). In V1 (D–F, far left columns), adaptation was strongest when adaptor and test stimuli corresponded to the preferred orientation of the neuron and decreased when adapting and testing with less preferred orientations (significant main effect of relative orientation, p = 4e-11). This relationship was particularly strong in neurons exhibiting high orientation selectivity (significant interaction between relative adaptor/test orientation and orientation selectivity, p = 0.005; for definition of orientation selectivity see above, Materials and Methods). Nevertheless, there was clear adaptation for all adaptor orientations as indicated by 1-back adaptation ratios consistently smaller than one (all p values < 0.004, corrected for multiple comparisons), except for nonpreferred (90°) adaptor and test stimuli of highly selective units (F, far left column, 90°, p = 0.88). This overall pattern of adaptation effects was qualitatively similar across cortical visual areas (D–F, columns 2–5). In thalamic areas (D–F, two far right columns), there was no evidence for a dependence of adaptation on orientation preference (no significant main effects of relative adaptor/test orientation: LGN, p = 0.28; LP, p = 0.91; no significant interactions between relative adaptor/test orientation and orientation selectivity: LGN, p = 0.24; LP, p = 0.92), likely because of the overall lower degree of orientation selectivity of thalamic neurons.

Figure 3.

Visual cortex, but not thalamus, exhibits long-term adaptation. A, Adaptation ratios of neurons in V1 as a function of the n-back trial. Strongest adaptation occurred in response to the 1-back stimulus, but stimuli encountered up to eight presentations in the past (seen 22 s ago) still exerted significant adaptation effect on the current visual response, despite the presentation of intervening stimuli (red bars, p < 0.05, corrected for multiple comparisons). The decay of adaptation over n-back trials was well captured by a double-exponential decay model with a fast- and slow-decaying adaptation component (black dashed line; a_fast = 13.99%, τ_fast = 0.85 trials, a_slow = 3.45%, τ_slow = 6.82 trials). Error bars denote bootstrapped 95% confidence intervals. B, Adaptation ratio as function of n-back trial for different visual areas (color coded). Although adaptation decays similarly and slowly across cortical visual areas (square symbols) and are generally significant for up to 6–8 trials back (symbols with black border, p < 0.05, corrected for multiple comparisons per area), it decays more rapidly in thalamic areas LGN and LP (circles). Black and lilac/green lines illustrate the best fitting exponential decay models for cortex and thalamus. Error bars indicate SEM. C, Average firing rates per mouse when the 4- to 8-back orientation was repeated (x-axis) or orthogonal (y-axis) relative to the current orientation. Mice exhibit consistent long-term adaptation in cortex (right) but not in thalamus (left). D, Histograms of single-neuron adaptation ratios (log transformed) in thalamus (left) and cortex (right).

Figure 4.

Cumulative adaptation effects in V1. Random sequences of grating orientations, as the ones used in the current experiment, prevent any systematic accumulation of adaptation across multiple stimulus presentations. Although this allows us to study the influence of individual n-back stimuli on the current visual response, it underestimates the influence of long-term adaptation in natural environments, in which orientations tend to remain stable over prolonged time periods (van Bergen and Jehee, 2019), therefore leading to an accumulation of adaptation. A, Illustrates that the adaptation effects of 2- to 8-back stimuli (red bars), albeit small when taken individually, together may lead to a considerable reduction of the current response (19% reduction, red-striped bar) that even outweighs the adaptation effect of the 1-back stimulus (17% reduction, light red bar). Importantly, the cumulative influence of repeating 2- to 8-back grating orientations could not be estimated empirically in the current dataset as such streaks of orientation repetitions are exceedingly rare for random sequences (probability of ∼0.006%). Here, we inferred the cumulative response reduction by assuming that the adaptation effects of previous stimuli accumulate approximately linearly. The inferred cumulative adaptation ratio was then calculated as $a{r}_{2-8}={\displaystyle \prod }_{n=2}^{8}a{r}_n,$ where ar_2-8 is the cumulative adaptation ratio of 2- to 8-back stimuli, and ar_n denotes the empirically estimated adaptation ratio of an individual n-back stimulus. B, To evaluate whether the assumption of a linear accumulation of adaptation approximately holds, we compared the empirically observed adaptation effect when two previous adjacent stimuli had the same orientation as the current stimulus (dark gray bars, ∼6.25% of all trials) to the cumulative adaptation effect inferred from individual n-back adaptation estimates (light gray bars). The empirically observed adaptation effect of two successive stimuli roughly matched the predicted adaptation effect, suggesting that adaptation accumulates approximately linearly in the current setting. All error bars indicate 95% CIs.

Figure 5.

Cortical long-term adaptation is driven by repeated stimulus orientations. We expressed the response modulation of neurons across all cortical areas by n-back repeated and orthogonal trials relative to a neutral baseline in which no stimulus was presented on the n-back trial. To this end, we computed adaptation ratios by dividing the firing rate of each neuron for repeat stimulus presentations by that of blank stimulus presentations (blue data points) or orthogonal divided by blank stimulus presentations (red data points). Although the suppressive effects of orthogonal stimuli decay quickly, repeated stimuli exert long-term suppression for up to eight trials. Error bars indicate bootstrapped 95% confidence intervals.

Figure 6.

Visual cortex exhibits adaptation in response to immediately preceding briefly presented static gratings. A, Presentation sequence of static grating stimuli. Mice were shown static gratings with a duration of 250 ms with no intervening gray period. Gratings had one of six orientations (0, 30, 60, 90, 120, 150°), five spatial frequencies (0.02, 0.04, 0.08, 0.16, 0.32 cycles/degree), and four phases (0, 0.25, 0.5, 0.75). The order of grating presentations was randomized. Similar to the analysis of drifting gratings, we contrasted activity to gratings preceded by gratings of the same orientation (repeat, blue) with that elicited by gratings preceded by a grating of the orthogonal orientation (orthogonal, red). B, Population peristimulus time histograms of neurons in V1 for repeat and orthogonal conditions. The visual response to the current stimulus (green shaded area) was reduced when the previous stimulus had the same orientation as the current stimulus (repeat), indicating orientation-specific adaptation. Vertical dashed lines denote onset and offset of the current stimulus, respectively. Bin width = 25 ms. Error bars indicate SEM. C, One-back adaptation ratios across visual areas. All areas show significant 1-back adaptation. Error bars indicate bootstrapped 95% confidence intervals. White numbers indicate the number of neurons in each area. D, Mice show consistently reduced firing rates after a repeated versus orthogonal orientation, as indicated by data points falling above the diagonal. Same conventions as in Figure 1E. E, Histograms of single-neuron adaptation ratios (log transformed) in thalamus (left) and cortex (right).

Figure 7.

Visual cortex exhibits long-term adaptation following briefly presented gratings. A, Adaptation ratios of V1 as a function of the n-back trial. Although adaptation was most strongly driven by the previous stimulus (1-back), stimuli encountered up to 20 presentations in the past (5 s ago) still exerted significant adaptation effects on the current visual response (red bars, p < 0.05, FDR corrected). Similar to drifting grating adaptation, the decay of adaptation over n-back trials was well captured by a double-exponential decay model with a fast- and slow-decaying adaptation component (black dashed line; a_fast = 8.17%, τ_fast = 0.54 trials, a_slow = 2.04%, τ_slow = 9.12 trials). Error bars indicate bootstrapped 95% confidence intervals. B, Adaptation ratios as a function of n-back trial for different visual areas (color coded). In cortical areas (squares) there is significant adaptation to stimulus orientations presented up to 20 trials back (symbols with black border, p < 0.05, FDR corrected per area), whereas in thalamic areas (circles) long-term adaptation is less evident. Error bars indicate SEM. Black and orange/green lines denote the best fitting exponential decay models for cortex and thalamus, respectively. Adaptation was computed over the whole stimulus interval (0–250 ms) because with the back-to-back presentation of static gratings, visual responses to the previous stimulus overlapped with the initial time window of the current stimulus, thereby increasing response variability in this early time window. However, largely similar results were obtained when performing the analyses on the same time window used in the drifting grating experiment (0–100 ms), except for a less clear difference of the decay of adaptation between cortex and thalamus. C, Average firing rates per mouse when the 5- to 20-back orientation was repeated (x-axis) or orthogonal (y-axis) relative to the current orientation in the thalamus (left) and cortex (right). D, Histograms of single-neuron long-term (average 5–20 back) adaptation ratios (log transformed) in thalamus (left) and cortex (right).

Figure 8.

Short-term (1-back) adaptation does not introduce spurious long-term adaptation effects for the particular stimulus sequences used in the experiments. A, B, We simulated responses of an artificial neuron to the particular stimulus sequences used in the drifting grating experiment (A) and static grating experiment (B). The artificial neuron responded equally to all stimulus orientations, but selectively reduced its responses to a successive repeated orientation to mimic orientation-specific 1-back adaptation. We chose the strength of this 1-back adaptation effect to match the empirically observed 1-back adaptation of V1. We subsequently analyzed the simulated responses with the same procedure used for the empirical data. The analysis of the simulated responses recovered the ground truth 1-back adaptation effect (black data points). There were no spurious adaptation effects for stimuli farther in the past, as indicated by the black data points being centered on an adaptation ratio of one, markedly different from the empirically observed long-term adaptation effects (red data points, adaptation in V1). Black error bars indicate 95% CIs of adaptation across the simulations of the 32 stimulus sequences. Red error bars indicate 95% CIs of empirical adaptation across neurons in V1.