Activity recall in a visual cortical ensemble.

Cue-triggered recall of learned temporal sequences is an important cognitive function that has been attributed to higher brain areas. Here recordings in both anesthetized and awake rats demonstrate that after repeated stimulation with a moving spot that evoked sequential firing of an ensemble of primary visual cortex (V1) neurons, just a brief flash at the starting point of the motion path was sufficient to evoke a sequential firing pattern that reproduced the activation order evoked by the moving spot. The speed of recalled spike sequences may reflect the internal dynamics of the network rather than the motion speed. In awake rats, such recall was observed during a synchronized ('quiet wakeful') brain state having large-amplitude, low-frequency local field potential (LFP) but not in a desynchronized ('active') state having low-amplitude, high-frequency LFP. Such conditioning-enhanced, cue-evoked sequential spiking of a V1 ensemble may contribute to experience-based perceptual inference in a brain state-dependent manner.

An essential feature of memory recall is pattern completion, in which the retrieval of memory is triggered by a subset of the cues present during learning. In particular, cue-triggered recall of a learned temporal sequence is useful for predicting future events based on current sensory inputs[1, 2]. Studies of episodic memory in human subjects have strongly indicated the role of the hippocampus and its surrounding cortical areas in the learning and recall of temporal sequences[3, 4]. Experiments in the rat also showed that hippocampal lesions can cause selective impairments of the ability to remember the temporal sequence of recently experienced sensory stimuli or spatial locations[5-7]. Furthermore, simultaneous recordings from a large number of hippocampal neurons have revealed spontaneous reactivation of the spike sequences observed during active exploration in subsequent periods of sleep or awake immobility[8-13]. These findings have pointed to the experience-dependent enhancement of sequential spiking in the hippocampus as a potential neural mechanism for sequence learning and memory. However, whether learning of simple sequences can also occur in early sensory circuits remains unclear.

Experience-dependent modifications in early sensory circuits are believed to contribute to some forms of perceptual learning. In the somatosensory[14] and auditory[15] cortex, learning causes an increased cortical representation of the learned stimuli. In the human visual system, the location, orientation, or spatial frequency specificity of training-induced improvement in pattern discrimination suggests an involvement of early visual circuits[16, 17]. Functional MRI studies further showed that learning in visual detection is accompanied by an increase in V1 activity[18]. Electrophysiological studies in awake monkeys have shown that training in orientation discrimination causes a change in the tuning curves of V1 neurons[19], and learning in shape discrimination causes changes in the contextual modulation of V1 neuronal responses[20]. Although these studies provided ample evidence for learning-related modification in adult V1, change related to sequence learning has not been demonstrated.

In this study, we tested learning and recall of a simple motion sequence in V1. Since V1 neurons have well-defined receptive fields (RFs), we used a moving spot to evoke sequential firing of the ensemble of neurons whose RFs fall along the motion path. Shortly after repeated motion conditioning, a briefly flashed spot at the starting point of the motion path evoked more sequential spiking similar to that evoked by the moving spot, reminiscent of cue-triggered recall of a learned temporal sequence. This recall effect was specific to the trajectory of the repeated motion, and in awake animals it depended on the brain state. Thus, cue-triggered recall of a simple spike sequence can occur in an early sensory circuit and may contribute to experience-based perceptual inference.

RESULTS

A multielectrode array with 2 × 8 channels was used to record the spiking activity of a neuronal ensemble in V1 of either urethane-anesthetized (Supplementary Fig. 1a) or awake head-fixed rat (Fig. 1a and Supplementary Movie 1; see Online Methods). The multiunit RFs mapped with sparse noise stimuli were distributed in a ~100 × 25 region of the visual field (Fig. 1b, c and Supplementary Fig. 1b, c), in a manner consistent with the retinotopic map of rat V1. Although eye movement was observed in the awake head-fixed animals, the amplitudes were much smaller than the average RF size (s.d. < 10% RF size, Supplementary Fig. 2), consistent with previous studies[21, 22]. Accordingly, the RFs in the awake rats were highly stable over time (Fig. 1b) and similar in size to those measured in anesthetized animals (Supplementary Fig. 1b, c), whose eyes were stabilized mechanically (see Online Methods). To evoke sequential spiking of the recorded neurons, we presented a conditioning stimulus consisting of a bright spot (anesthetized: 13.5°; awake: 18° in diameter) moving repeatedly from S (starting point) to E (end point), across the long axis of the RF distribution (speed 180° s−1, repeated once every 2.1 – 2.3 s, 100 trials; Fig. 1c and Supplementary Fig. 1c). In both anesthetized and awake rats, this conditioning stimulus evoked strong spiking responses with the firing sequence largely consistent with the RF locations along the motion trajectory (Fig. 1d and Supplementary Fig. 1d). The Spearman (rank-order) correlation coefficients (CCs) between the firing time of the recorded units and their RF locations along the S → E axis were 0.94 ± 0.08 (awake, mean ± s.d., n = 18 experiments) and 0.91 ± 0.09 (anesthetized, n = 19 experiments).

Figure 1

Sequential spiking of neuronal ensemble in rat V1 evoked by moving spot. (a) Schematic of experimental setup. Visual stimuli were presented to left eye, multielectrode array (red dots) was inserted into right V1. Diagram of rat cortex was adopted from Paxinos and Watson[48], shown on 1 × 1 mm grid. A, anterior; P, posterior; L, lateral. V1, primary visual cortex. V1M, monocular region of V1; V1B, binocular region of V1; V2M, medial secondary visual cortex; V2L, lateral secondary visual cortex. λ, skull landmark lambda. (b) Multiunit RFs recorded simultaneously by 16 electrodes in an awake rat. Red ellipse, contour of Gaussian fit at one standard deviation. The two repeats of RF measurement were separated by 30 min. “*”, units excluded from analyses because of low firing rate or extraneous RF position (see Online Methods). (c) Superposition of RFs and visual stimuli. Colored ellipses, Gaussian fits of RFs. White dashed circles, “Starting point” (S) and “End point” (E) of conditioning spot. (d) PSTHs of the 14 units shown in (c) during conditioning, ordered by distance between RF center and S. Red curves, PSTHs smoothed with Bayesian adaptive regression splines.

Conditioning-induced increase in sequential spiking

To assess learning of this motion sequence by the cortical ensemble, we briefly presented the bright spot at the starting point of the motion path (S) as the test cue and examined the spike sequence within the cortical ensemble. Figure 2a, b shows the spike trains of simultaneously recorded multiunits evoked by the test cue at S in several consecutive trials before (top row) and after (bottom row) 100 repeats of the motion conditioning (middle row), with the units ordered by their RF locations along the S → E axis. The test cue-evoked spike trains in Figure 2b are shown at an expanded time scale, since the recall in awake animals was found to occur at a faster speed (see below). Prior to the conditioning, the responses in some test trials exhibited sequential spiking, but the firing sequence was quite variable across trials. In the trials after conditioning, we found a marked increase in sequential spiking similar to that evoked by the conditioning motion stimuli.

Figure 2

Conditioning-induced increase of sequential spiking in anesthetized and awake rats. (a,b) Example spike trains of simultaneously recorded units in anesthetized (a, 13 units) and awake (b, 12 units) rats evoked by S → E conditioning (“cond”; middle) and by test cue at S before (top) and after (bottom) 100 trials of conditioning at a speed of 180° s−1. Each row shows responses in 7 consecutive trials, 0 – 800 ms (a, middle row of b) or 0 – 250 ms (b, top and bottom rows) after stimulus onset. Units are ordered by distance between RF center and S. (c,d) Top three rows, pairwise cross-correlation averaged from anesthetized (c, n = 19 experiments) and awake (d, n = 18) rats with 100 trials of conditioning at 180° s−1. Bottom row, difference between cross-correlation functions before and after conditioning.

To further demonstrate such sequential spiking, we adopted a pairwise correlation analysis previously used to characterize spontaneous sequence replay in the hippocampus and prefrontal cortex[11, 23]. Briefly, we computed the cross-correlation between each pair of spike trains for all pairwise combinations of simultaneously recorded units in each experiment; the correlation was then plotted against both the temporal delay between the pair and their RF distance. For the responses to conditioning motion, the peak of correlation showed a prominent rightward shift with the RF distance in both anesthetized and awake animals (upper row; Fig. 2c, d). This lower left to upper right slant of the correlation function represents an increase in the relative spike timing of the pair with their RF distance, thus serving as a reliable indicator for sequential spiking within the ensemble. For the responses evoked by the test cue at S, there was a slight slant of the correlation function before conditioning (second row; Fig. 2c, d). This is because the response to a flashed stimulus generally exhibits some horizontal spread within the cortex[24]. Although the response to the test cue at S is expected to spread symmetrically from S (both towards and away from E), the recorded neurons were all on one side of S (towards E). The slight slant in the correlation function among these units thus reflects the spread of the response. Notably, there was a marked increase in the slant after conditioning (third row; Fig. 2c, d). The conditioning-induced change is also shown by the difference function (“After – Before”, bottom row), which revealed an increase in correlation at positive timing and a decrease at negative timing with a similar slant. This indicates that the motion conditioning enhanced test cue-evoked sequential spiking of the cortical ensemble in the S → E direction. This effect was not accompanied by any significant change in the mean firing rates of the neurons (anesthetized, P = 0.80, n = 18 experiments; awake, P = 0.84, n = 19; Wilcoxon signed rank test), indicating that the effect was due to changes in the relative timing rather than the rate of neuronal spiking. Such a conditioning-induced modification of the cue-evoked ensemble spike pattern provides a potential neural mechanism for learning and recall of the motion sequence.

Specificity of sequence recall

To analyze the effect of conditioning on sequential spiking quantitatively, we computed for each test trial the Spearman CC between the firing time of the units and their RF locations along the S → E axis (see Online Methods). As shown in the cumulative histograms of CCs (Fig. 3a, b), 100 trials of motion conditioning at a speed of 180° s−1 caused a significant rightward shift of the CC distribution in both anesthetized (mean CC from 0.21 to 0.29, P = 1.5 × 10−3, n = 19 experiments; Kolmogorov-Smirnov test) and awake (mean CC from 0.22 to 0.30, P = 1.5 × 10−4, n = 18) animals, indicating increased sequential firing along the S → E trajectory. Instead of correlation between the firing time and RF location, we also computed the Spearman CC between the neuronal firing time for S-evoked responses and that for conditioning stimulus-evoked responses. We found a similar rightward shift of the CC distribution after conditioning (Supplementary Fig. 3).

Figure 3

Specificity of cue-triggered recall of spike sequence. (a,b) Cumulative histograms of Spearman CCs between test cue-evoked spike sequence and RF position, before (dotted line) and the first 2 min after (solid line) 100 trials of conditioning at 180° s−1 for anesthetized (a, n = 19 experiments) and awake (b, n = 18) rats. Note that the CC distribution before conditioning (dotted line) shows a rightward shift from CC = 0, due to symmetric spread of activity evoked by the stimulus at S. (c,d) Cumulative histograms of CCs for spike sequences evoked by test cues at mid-point between S and E (c, n = 19) and at E (d, n = 18) before (dotted line) and the first 2 min after (solid line) conditioning in awake rats. (e,f) Cumulative histograms of CCs for test stimuli at S (e) and E (f) before (dotted line) and the first 2 min after (solid line) 100 trials of flashed bar conditioning (n = 18). Since there is no directionality in the bar stimulus, the starting and end points were chosen randomly for each experiment as either of the two ends of the bar, referred to as pseudo-S (e) and pseudo-E (f), respectively. Diagram in upper left region of each plot illustrates conditioning and test stimuli used in that experiment.

Note that the CC distribution for S-evoked responses showed a rightward bias even before conditioning (dotted line; Fig. 3a, b). As explained above, this reflects the symmetric spread of the activity evoked by the test cue at S. To eliminate this bias, we performed additional experiments in awake rats with the test spot flashed at the mid-point between S and E (see Online Methods). As expected, we found no rightward bias in the CC distribution before conditioning (dotted line; Fig. 3c), as the RFs of the recorded units were distributed on both sides of the test spot. Nevertheless, 100 trials of S → E conditioning still caused a significant rightward shift of the CC distribution (mean CC from −0.08 to −0.02, P = 1.3 × 10−4, n = 19 experiments; Kolmogorov-Smirnov test; solid line; Fig. 3c). Furthermore, to test whether the observed recall was specific to the conditioned motion sequence (S → E), we analyzed the response to the test stimulus presented at the end point (E), which in principle could evoke an E → S spike sequence. Before conditioning, the CC between the timing of E-evoked spiking and the RF position along the E → S axis also showed a rightward bias (dotted line; Fig. 3d and Supplementary Fig. 4a) that was not significantly different from that for S-evoked spiking (dotted line; Fig. 3a, b; awake: P = 0.58, n = 18 experiments; anesthetized: P = 0.26, n = 19; Kolmogorov-Smirnov test), again reflecting the symmetric spread of cortical response. However, unlike for the S-evoked responses, we found no significant change in the CC distribution for the E-evoked spiking after S → E conditioning (awake: mean CC from 0.20 to 0.20, P = 0.59, n = 18 experiments; anesthetized: mean CC from 0.20 to 0.21, P = 0.77, n = 19; Kolmogorov-Smirnov test; Fig. 3d and Supplementary Fig. 4a). Thus, instead of a non-specific increase in the horizontal spread of the cortical response to a flashed stimulus, S → E conditioning caused a selective increase in sequential firing in the S → E direction. Furthermore, we found that 100 trials of conditioning along a motion path parallel to, but not overlapping with, the long axis of the recorded RF distribution (at a distance of 27° – 45°) caused no significant change in the CC distribution for either S- or E-evoked spiking (S: mean CC from 0.37 to 0.36, P = 0.80, n = 14; E: mean CC from 0.42 to 0.40, P = 0.93; Kolmogorov-Smirnov test; Supplementary Fig. 5), indicating that the effect was also specific to the position of the motion sequence rather than reflecting a general change in the direction of spike propagation in the visual cortex.

To further confirm the importance of the sequential excitation evoked by the S → E conditioning, we used a flashed bar spanning the region between S and E as the conditioning stimulus in awake rats (see Online Methods). We found that 100 trials of the bar conditioning caused no significant change in the CC distribution for the test cue at S (mean CC from 0.22 to 0.22, P = 0.76, n = 18 experiments; Kolmogorov-Smirnov test; Fig. 3e) or E (mean CC from 0.19 to 0.19, P = 0.62; Fig. 3f), although they caused a small but significant increase in the mean firing rates of the neurons (from 10.5 ± 0.1 to 11.0 ± 0.3, mean ± s.e.m., P = 0.03, Wilcoxon signed rank test). A previous study using voltage-sensitive dye imaging (VSDI) showed that following repeated flashes of a bright stimulus at a given retinal location, the spatiotemporal patterns of spontaneous cortical waves became more similar to the cortical response evoked by the flashed stimulus[25]. We thus tested whether such a flashed conditioning stimulus at S can cause an increase in sequential firing. As shown in Supplementary Figure 4b, c, 100 flashes of the bright spot caused no significant change in CC distribution for the test cue at S (mean CC from 0.30 to 0.31, P = 0.47, n = 17 experiments; Kolmogorov-Smirnov test) or E (mean CC from 0.32 to 0.32, P = 0.44). Together, these experiments indicate that the static conditioning stimuli are not sufficient to cause the increase in sequential firing.

In the experiments above, the electrode array was always implanted in V1 at an orientation similar to that illustrated in Figure 1a. To test whether the effect was specific to this cortical axis, we performed another set of experiments with the electrode array implanted at a new orientation (Supplementary Fig. 6a), with corresponding changes in the RF distribution (Supplementary Fig. 6b). We found that S → E conditioning also caused a marked increase in the slant of the pairwise cross-correlation function (Supplementary Fig. 6c) and a significant rightward shift of the CC distribution for the test cue at S (mean CC from 0.28 to 0.37, P = 4.3 × 10−5, n = 18 experiments; Kolmogorov-Smirnov test; Supplementary Fig. 6d), but not at E (mean CC from 0.31 to 0.29, P = 0.53; Supplementary Fig. 6e). Thus, the increase in sequential firing induced by the motion conditioning was not restricted to a particular cortical axis.

Further characterization of sequence recall

To facilitate further characterization of the effect of conditioning, we defined a ‘match’ as a test trial whose Spearman CC is above a given threshold and compared the percentage of matches before and after conditioning in each experiment. We found that the percentage increased significantly after conditioning over a wide range of CC thresholds for both awake and anesthetized animals for test cue at S (black line; Fig. 4a and Supplementary Fig. 7a), but not at E (gray line). To examine the persistence of the effect, we plotted the increase in the percentage of matches (at CC threshold of 0.6) measured at different times after 100 trials of conditioning. Within a period of ~7 min, the effect decayed completely for awake (Fig. 4b, c) and partially for anesthetized (Supplementary Fig. 7b, c) animals. We then tested whether the magnitude and persistence of the effect depend on the number of conditioning trials in anesthetized rats. As shown in Figure 5a, significant effect was found after 50, 100, and 200 trials of conditioning, but not after 10 trials. Although the initial magnitude of the effect was comparable among 50, 100, and 200 trials, the persistence of the effect appeared to increase with the trial number.

Figure 4

Persistence of conditioning-induced increase in sequential spiking in awake rats. (a) Difference in percentage of sequence matches before and after conditioning vs. CC threshold for test cues at S (black) and E (gray). The difference was significant at all CC thresholds below 0.75 for S (“*”, P < 0.05; “**”, P < 0.01; “***”, P < 0.001; Wilcoxon signed rank test), but not at any threshold for E. Error bar, s.e.m. (b) Time course for decay of conditioning-induced increase in percentage of matches (at CC threshold of 0.6). For S, the increase was significant at 2 min (P = 8.6 × 10−4; Wilcoxon signed rank test) and 4 min (P = 0.022), but not at 6 min (P = 0.57). For E, the effect was not significant at any time. Error bar, s.e.m. (c) Cumulative histograms of Spearman CCs at different periods after conditioning (solid lines). Dotted line, histogram before conditioning. The difference was significant at 2 and 4 min (P = 1.5 × 10−4; P = 9.8 × 10−3; Kolmogorov-Smirnov test), but not at 6 min (P = 0.1).

Figure 5

Dependence of sequence learning on conditioning parameters and NMDA receptor activation, measured in anesthetized rats. (a) Dependence of conditioning-induced increase in sequential firing on the number of conditioning trials. Each line represents decay time course of the increase in percentage of matches (at CC threshold of 0.6) induced by a given number of conditioning trials at a speed of 180° s−1. Red: 200 trials, P = 0.002, 0.01, 0.002 at 2, 4, 6 min after conditioning, n = 15; black: 100 trials, P = 3.3 × 10−3, 0.044, 0.027, n = 19; magenta: 50 trials, P = 0.03, 0.30, 0.42, n = 15; cyan: 10 trials, P = 0.90, 0.39, 0.87, n = 14; Wilcoxon signed rank test. (b) Dependence on conditioning speed. Each line represents decay time course following 100 trials of conditioning at a given conditioning speed. Dash: 360° s−1, P = 2.8 × 10−4, 0.01, 0.23, n = 18; Solid: 180° s−1 (same as black line in a); Dotted: 60° s−1, P = 0.64, 0.99, 0.41, n = 21. (c) Effect of local application of 75 μM APV. Black/gray line shows decay time course following 100 trials of conditioning at a speed of 180° s−1 with/without APV. Gray line, same as black solid line in (a) and (b). Black line, P = 0.34, 0.59, 0.18, n = 18. Error bar, s.e.m.

We also tested the dependence of the effect on the speed of conditioning motion. We found that 100 trials of conditioning at 180° s−1 and 360° s−1 caused similar increases in sequential firing, although conditioning at 60° s−1 was ineffective (Fig. 5b). We next asked whether the speed of the recalled spike sequence depends on the speed of conditioning motion. The speed of each recalled sequence was measured by linear regression of the RF locations of the recorded units as a function of their firing time. We found that in anesthetized rats the recall speed was similar for conditioning at 180° s−1 and 360° s−1 (~200° s−1, corresponding to ~5 mm s−1), which is slower than but within the same order of magnitude as the spontaneous wave speed observed with VSDI[25]. On the other hand, at the same speed of conditioning (180° s−1), it was much faster in awake than in anesthetized rats (Supplementary Fig. 8a). This is likely due to the difference in the intrinsic network dynamics: While the cortical LFP showed prominent slow oscillations (0.5 – 1 Hz) under urethane anesthesia, it exhibited faster dynamics in the awake animals (Supplementary Fig. 8b). Together, these results indicate that the speed of the recalled spike sequence is determined primarily by the network dynamics rather than by the speed of conditioning motion. Thus only the sequence, but not the speed, of the learned motion is represented in the recall. In spontaneous memory reactivation in the hippocampus and prefrontal cortex, the spike sequences observed during behavior are often replayed at a compressed time scale[9, 10, 23]. Interestingly, the duration of the replayed spike sequence is also determined by the duration of the sharp wave ripple rather than reflecting the highly irregular behavior of the rat on the running track, suggesting that the sequence rather than time course of the experience is re-activated[26].

In principle, the effect we have observed in V1 could reflect circuit modifications in other brain areas (e.g., retina and thalamus). To test whether the effect requires synaptic modifications within the visual cortex, we repeated the motion conditioning experiment after local application of D(−)-2-Amino-5-phosphonopentanoic acid (D-APV), a specific NMDA receptor blocker known to block spike timing-dependent plasticity (STDP)[27, 28] in the visual cortex[29]. We found that 75 μM of APV largely blocked the conditioning-induced increase in sequential firing (Fig. 5c). However, consistent with previous studies[30], blocking NMDA receptors also led to a significant reduction in cortical firing rate (Supplementary Fig. 9), leaving open the possibility that the blockade of the conditioning effect was caused by a general reduction in cortical spiking rather than by the specific block of NMDA receptor-dependent STDP. In addition, after topical application in V1, APV may also affect the neighboring cortical area V2 (see Online Methods). Nevertheless, this result strongly suggests that the conditioning effect depends on activity-dependent modification within the visual cortex rather than in earlier stages of the visual pathway.

Dependence on brain state

Some of the awake animals in our study exhibited frequent spontaneous switching between two behavioral states: a quiet immobile state with high-amplitude, low-frequency (< 10 Hz) LFP activity, here referred to as ‘synchronized/quiet’ brain state (distinct from slow-wave sleep by its lack of UP-DOWN state oscillations), and a ‘desynchronized/active’ state with facial/whisker movement and irregular, high-frequency LFP activity[31, 32] (Fig. 6a and Supplementary Movie 1). Since cortical neuronal excitability and synaptic efficacy are differentially influenced by neuromodulators under different brain states, we inquired whether the observed sequence recall depends on the brain state.

Figure 6

Dependence of sequence recall on brain state in awake rats. (a) Brain state switch indicated by changes in LFP power spectrum. Red and blue lines, periods of synchronized (“Synch”) and desynchronized (“Desynch”) states, corresponding to red and blue clusters in (b). Insets above, sample LFP traces during periods marked by red and blue blocks. (b) Clusters in 2-D state space (Ratio 1, ratio of LFP powers between 1 – 10 Hz and 1 – 25 Hz; Ratio 2, between 15 – 30 Hz and 15 – 60 Hz). (c,d) Pairwise cross-correlations (similar to Fig. 2c, d) in synchronized (c) and desynchronized (d) states. (e,f) Cumulative histograms of Spearman CCs for test cue at S before (dotted line) and the first 5 min after (solid line) conditioning during synchronized (e) and desynchronized (f) states for the 7 experiments with frequent switching between the two brain states.

For 7 experiments in awake rats with frequent switching between the two brain states, we separated each recording into synchronized and desynchronized periods using cluster analysis of the LFP power spectrum (Fig. 6b; see Online Methods). The spike sequences evoked by the test cue at S were analyzed within each state and compared before and after conditioning. We found that for the test trials in the synchronized brain state, the pairwise cross-correlation function showed a strong increase in the slant after conditioning (Fig. 6c), and the CC distribution showed a marked rightward shift (mean CC from 0.22 to 0.37, P = 2.4 × 10−3, n = 7 experiments; Kolmogorov-Smirnov test; Fig. 6e). For the desynchronized state, however, conditioning did not cause an increased slant in the cross-correlation function (Fig. 6d) or a significant shift in the CC distribution (mean CC from 0.31 to 0.29, P = 0.43; Fig. 6f). To test whether this difference is due to different degrees of eye movement in the two brain states, we measured eye movement in each state with a video camera (see Online Methods). Although there was indeed a higher level of eye movement in the desynchronized state, the amplitude was much smaller than the RF size (s.d. < 10% of RF size; Supplementary Fig. 2) and thus unlikely to exert significant effects on the visual inputs. This indicates that although motion conditioning causes a significant change in the cortical circuit, manifestation of the change in the ensemble spiking activity depends strongly on the brain state.

DISCUSSION

Our study suggests that cue-triggered recall of a recently experienced temporal sequence, a function previously attributed to higher brain areas, could be partly supported by V1 ensemble activity. The observed increase in sequential firing may result from activity-dependent synaptic modification. The conditioning-evoked sequential spiking of neighboring neurons (Fig. 1d and Supplementary Fig. 1d) is well suited for the induction of STDP of intracortical connections, which should selectively potentiate the S → E connections and facilitate spike propagation in the same direction. This is reminiscent of Hebb’s notion of sequential activation of cell assemblies in a ‘phase sequence’[33]. Previous studies in V1 of anesthetized animals have demonstrated stimulus-induced RF modifications with temporal specificity comparable to STDP[34, 35]. The current study suggests that such synaptic plasticity may also modify neuronal ensemble dynamics to support sequence learning and pattern completion, consistent with theoretical predictions[36, 37].

Our study showed that a static stimulus flashed repeatedly at S was ineffective for enhancing sequential firing (Supplementary Fig. 4b, c), while a previous study using VSDI showed that a similar conditioning stimulus caused a significant increase in the similarity between the spontaneous and visually evoked cortical waves[25]. The difference between these findings may be attributed to multiple factors. First, while the VSDI signals reflect both supra- and subthreshold activity[38], the multielectrode recording used here measures only the spiking activity. Second, the increase in the similarity between the spontaneous and visually evoked waves observed with VSDI could be attributed at least partly to changes in the initiation sites of the spontaneous waves[25]. Whether the flashed conditioning stimulus induced any change in the direction of wave propagation is unclear. It is possible that while the flashed conditioning can cause an increase in the excitability of the stimulated cortical region, thus the shift in the initiation sites of spontaneous waves, motion stimuli are required for enhancing the horizontal connections, leading to the increase in cue-triggered sequential firing.

The notion that the observed sequence learning is mediated by a simple synaptic mechanism without requiring high-level cognitive processes is consistent with the presence of the effect in anesthetized animals and with elimination of the effect by local APV application (Fig. 5c), although the effect of APV could be mediated by the reduction of cortical spiking rather than by the block of NMDA receptor-dependent synaptic modifications. On the other hand, the dependence of the effect on brain state (Fig. 6) points to the susceptibility of sequence recall to top-down modulations. The fact that significant effect was observed in both anesthetized and quietly awake states but not in the desynchronized active state suggests that the cue-triggered recall of a learned sequence depends on correlated activity among a large number of cortical neurons. Furthermore, the desynchronized active brain state may be associated with enhanced thalamocortical inputs and weakened intracortical connections, due to the influences of neuromodulators such as acetylcholine[39, 40]. A reduced contribution of the intracortical connections that have been modified by the motion conditioning could diminish the conditioning effect. Note that in our experiments, the rightward shift of CC distribution induced by 100 trials of conditioning was larger in the synchronized awake state (Fig. 6e) than in the anesthetized state (Fig. 3a), suggesting that learning of the sequence also depends on the brain state. In future studies with experimental control of the brain state during conditioning, it would be interesting to examine how learning depends on the brain state[41, 42], and whether recall also depends on the brain state at which the learning occurs[43].

Functionally, perception depends on both the current sensory inputs and previous experience, and their relative importance may vary across behavioral states[44]. While the desynchronized state may favor faithful representation of sensory inputs, the correlated ensemble activity in the synchronized state may carry information about previous experience. Repeated occurrence of a motion sequence in the recent past may suggest higher probability of such a sequence in the near future. The cue-triggered recall of spike sequence may thus provide a powerful mechanism for experience-dependent perceptual inference[45, 46]. From this point of view, the relatively short persistence of the effect (on the order of minutes) may be functionally advantageous, since experience from the recent past should be a good predictor only of the events in the near future. On the other hand, with more repeats of the stimulus sequence, especially if spread out over days or weeks, the effect may become more persistent. In addition to perceptual inference shortly after the experience, the cue-triggered recall we have observed may also complement the spontaneous reactivation of recent experience found in a variety of neuronal circuits[8-12, 23, 25, 47] to facilitate long-term memory consolidation.