Learned spatiotemporal sequence recognition and prediction in primary visual cortex

Abstract

Learning to recognize and predict temporal sequences is fundamental to sensory perception and is impaired in several neuropsychiatric disorders, but little is known about where and how this occurs in the brain. We discovered that repeated presentations of a visual sequence over a course of days resulted in evoked response potentiation in mouse V1 that was highly specific for stimulus order and timing. Notably, after V1 was trained to recognize a sequence, cortical activity regenerated the full sequence even when individual stimulus elements were omitted. Our results advance the understanding of how the brain makes 'intelligent guesses' on the basis of limited information to form visual percepts and suggest that it is possible to study the mechanistic basis of this high-level cognitive ability by studying low-level sensory systems.

Fig. 1

Learned spatiotemporal sequence representations in V1. (a) Schematic representation of head–fixed stimulus presentation. (b) On each of four training days, the experimental group (n=6) was shown 200 presentations of the sequence ABCD (where each letter indicates a uniquely oriented sinusoidal grating) and the control group (n=4) was shown 200 random permutations of the same sequence elements. Each element was held onscreen for 150 ms and sequences were separated by 1.5 s of gray screen. All animals were tested on the 5th day with the sequences ABCD, DCBA, and ABCD₃₀₀ (subscript indicates a 300 ms element hold time). (c) Sequence evoked local field potentials recorded on the 5th day show that ABCD drives larger responses (blue) than DCBA (red) in the experimental animals, while there is no differential responses in control animals. Voltage traces represent the average response of all animals in each group and triangles mark the onset of each sequence element. (d) ABCD₃₀₀ drives relatively small responses in both groups. (e) Training regime has a significant effect on sequence response magnitude (quantified as the average peak–to–peak response to each of the four elements, see Supplementary Fig. 1) potentiation (2–way RM ANOVA, F_(1,8)=22.560, P=0.001). There is a significant interaction (2–way RM–ANOVA, F_(1,8)=6.638, P=0.008) between sequence and experimental group on day 5 and post–hoc analysis shows that the response to ABCD is significantly larger than either DCBA (t₍₅₎=5.738, P=0.002) or ABCD₃₀₀ (t₍₅₎=4.923, P=0.005). Sequence effects are not significant within the control group. Error bars show s.e.m. (f) Potentiation time course. (g) Sequence effects are evident in spiking neural activity. In this representative example, ABCD drives higher peak firing rates than DCBA (multi–unit spike rasters above peristimulus time histograms, dashed lines indicate element onset times).

Fig. 2

Sequence learning is temporally specific. (a) Mice (n=13) were trained using ABCD presented with a short–long–short–long temporal profile. On the fifth day, the mice were tested with ABCD and DCBA presented with both familiar (black) and novel (long–short–long–short, gray) timing. (b) The largest responses occur when the trained sequence is presented with the trained timing (top). Timing makes little apparent difference when a novel sequence is shown (bottom). (c) There is a significant interaction between sequence order and timing (2–way RM ANOVA, F_(1,12)=22.925, P<0.001). Post–hoc analysis shows the response to ABCD with trained timing is significantly larger than ABCD with novel timing (t₍₁₂₎=8.760, P<0.001). There is also a small effect of timing within DCBA (t₍₁₂₎=2.722, P=0.012). Error bars show s.e.m. (d) The relative effect of timing as a function of sequence is demonstrated by paired–response plots (dashed lines connect responses for single animals, black indicates the mean).

Fig. 3

Learning does not transfer between eyes and requires muscarinic acetylcholine receptors in V1 but not NMDA receptors. (a) Mice (n=8) were trained with an occluder restricting visual stimulation to the left eye (LE). Responses were recorded in the hemispheres contralateral and ipsilateral to the viewing eye. On the 5th day, sequences were presented to both eyes. (b) ABCD drives larger responses than DCBA in both hemispheres only when viewed through the trained eye. (c) There is a significant interaction between viewing eye and sequence in both hemispheres (2–way RM ANOVA, Contra: F_(1,7)=25.041, P<0.001 Ipsi: F_(1,7)=10.426, P=0.002). The response to ABCD is significantly larger than DCBA in both hemispheres only when viewed through the trained eye (Contra: t₍₇₎= 8.246, P<0.001, Ipsi: t₍₇₎=5.091, P<0.001). (d) Systemic CPP treatment (left, n=9, 30–60 min prior to stimulus presentation during training) has no significant effect on sequence potentiation compared to vehicle (n=6) and response potentiation is significant within both treatment groups (main effect: F_(1,13)=35.525, P<0.001, post–hoc analysis: t₍₈₎=3.186, P=0.007 and t₍₅₎=5.093, P<0.001). The same CPP blocked subsequent SRP induction in the same mice (right, regrouped after washout, CPP n=6, vehicle n=9). There is a significant interaction between treatment and SRP recording session (2–way RM–ANOVA, F_(1,13)=42.210, P<0.001) and potentiation is significant only in the vehicle control group (t₍₈₎=9.692, P<0.001). (e) Muscarinic receptor antagonism during training blocks sequence potentiation. There is a significant day 5 interaction between treatment and sequence in scopolamine (n=5) and vehicle (n=5) treated mice (2–way RM ANOVA, F_(1,8)=5.827, P=0.013) and ABCD is significantly larger than DCBA (t₍₄₎=3.661, P=0.004) or ABCD₃₀₀ (t₍₄₎=3.813, P=0.005) only in vehicle treated mice. (f) Local unilateral infusion of scopolamine in V1 (n=7 mice) blocks potentiation relative to the opposite vehicle treated hemisphere (2–way RM ANOVA, F_(1,6)=30.189, P=0.002). Error bars show s.e.m.

Fig. 4

Learned sequence representations are predictive and involve multiple cortical layers. (a) Mice (n=7) were trained with ABCD and tested with two sequences, A_CD and E_CD, where the second element was replaced with a gray screen. (b) When the omitted element is preceded by A (red), a response occurs in position 2 (marked by the dashed gray box) that is similar in form and latency to the response evoked when B is actually presented (blue). This predictive response is absent when the omitted element is preceded by the novel element E (green). (c) There is a significant effect of sequence on both the average magnitude across the four elements (left, 1–way RM ANOVA, F_(2,6)=12.186, P=0.001) and the response of the second element alone (right, 1–way RM ANOVA, F_(2,6)=31.597, P<0.001). Significant differences determined by post–hoc analysis are indicated by brackets (Full sequence: t₍₆₎=4.675, P=0.002 and t₍₆₎=3.711, P=0.006, Elmnt 2: t₍₆₎=4.175, P=0.003 and t₍₆₎=3.771, P=0.003). Error bars show s.e.m. Laminar field recordings (d) and CSD analysis (e) show characteristic activation patterns evoked by trained and novel sequences. The DCBA sink–source pattern is similar to ABCD but with smaller magnitudes. Activation patterns during omitted elements (marked gray triangles) closely match those produced by real stimuli when the sequence is initiated with A but not E. When each sequence element is held onscreen for twice the trained duration, activation patterns resembling those that would have occurred had element B been shown appear at the expected time (highlighted with dashed gray box).