Local/global oddball paradigm, experimental setup, and predictive routing.

(a) Neuropixels were introduced into 6 visual cortical regions in all mice (V1, LM, RL, AL, PM, AM). (b) How predictive routing models local and global oddballs: in the local oddball, a feedback ⍺/β oscillation predicting stimulus ‘x’ will fail to inhibit cortical columns responsive to stimulus ‘y’, resulting in uninhibited ɣ oscillations and spiking upon the onset of an unexpected ‘Y’, while in the global oddball, a feedback ⍺/β oscillation predicting stimulus ‘y’ after “xxx” will fail to inhibit cortical columns responsive to stimulus ‘x’, resulting in uninhibited ɣ oscillations and spiking upon the onset of the unexpected repetition ‘X’. (c) N=14 animals were habituated to the local oddball sequences (“xxxY” in Cohort 1 of 7 animals, “yyyX” in Cohort 2 of 7 animals) before recording; stimuli were presented for 500 ms with a 500-531 ms inter-stimulus interval and 1 s inter-trial intervals. (d) Visual stimuli were presented in 4-stimulus sequences consisting of two (2) oriented, drifting bar gratings (symbolized as ‘x’ or ‘y’), each displayed full-screen. After the main block of recording in which local and global oddballs were presented, the random control block consisted of randomized individual stimuli presented in sequences of four, while the repetition control block consisted of sequences of four repetitions of the same stimulus. (e) Local field potential in the middle channel of L4 over the course of the ‘y’ stimulus in the random deviant condition in the higher-order visual area PM; the plot shows oscillatory activity before, during, and after the stimulus-driven negative deflection in the local field potential. (f) Schematic of how surprise affects spectral responses according to hypotheses H1 (predictive routing), H2 (ɣ encoding prediction), and H3 (aperiodic shifts). Dotted lines show the aperiodic component of the spectral response, while solid lines show the reconstructed spectrum that includes oscillatory peaks. H1 (blue) predicts that surprise increases θ power, decreases ⍺/β power, and increases ɣ power; H2 (orange) predicts that surprise decreases ɣ power while confirmation increases it; and H3 (green) predicts that surprise shifts the aperiodic component of the spectrum. Created in BioRender. Sennesh, E. (2025) https://BioRender.com/ogllv24.

Probabilistically deviant stimuli induce a “push-pull” between

⍺/β- and ɣ oscillations. The top row shows the two conditions contrasted to examine deviance detection; the middle row shows the laminar signatures of the contrast results broken out by frequency band; and the bottom row shows the laminar CSD responses characteristic of each frequency band in these conditions. Solid contour lines enclose the area of a synchronization/power increase, while dotted contour lines enclose the area of a desynchronization/power decrease. (a) Spectral responses to a fully randomized “XXXY” (n=541 trials) conform to stereotyped “oddball” responses, with stimulus onset leading to an increase (2.5-5 dB) and then decrease (0-2.5 dB) in ⍺/β power over baseline, then to a long-lasting increase (2.5-5 dB) in ɣ (40-55 Hz) synchronization. (b) Spectral responses to a fully predictable, habituated “xxxy” (n=700 trials) show a fast increase (>5 dB) over baseline and then return to baseline of ⍺/β power, followed by a long-lasting increase (2.5-5 dB) in ɣ (40 rising to 40-55 Hz over time) synchronization. (c) Detection of probabilistic deviants, via the contrast between “XXXY” and “xxxy”, induces a 1 dB decrease in ⍺/β power and a 1-2 dB ɣ power increase. (d) Deviance-induced ɣ (30-90 Hz) synchronization occurs earliest and most strongly at the boundary between L1 and L2/3, remaining strong in L1 while spreading down into L2/3; a 0-1 dB loss of ɣ power occurs in L5/6 100ms into deviance detection. (e) Deviance-induced ⍺/β (10-30 Hz) activity starts with a fast-onset increase in synchronization in L1, followed by ⍺/β desynchronization (1.5-2 dB) in L/3 and down into L5, followed by a slow decrease in ⍺/β power (0.5 dB) in L5/6 and increase in ⍺/β power (1-1.5 dB) in L1. (f) Deviance detection induces a weak (0.5-1 dB), late increase in θ (2-10 Hz) power. (g) ɣ oscillations in P4 of the fully randomized “XXXY” sequence display a canonical feedforward pattern of laminar activation, beginning in L4, then flowing out to L/3 and later into L5/6, over n=125 complete oscillatory cycles re-epoched from 541 trials. Oscillatory LFP trace shown below for illustrative purposes only. (h) ⍺/β oscillations in P4 of the fully habituated “xxxy” sequence display a weak but canonical feedback pattern of laminar activation, beginning in L1 and flowing down to L2/3, then down through L4 to meet a rising flow of activation in L5. CSD was calculated over the average of n=189 complete oscillatory windows over 700 trials. Oscillatory LFP trace shown below for illustrative purposes only. (i) θ oscillations in P4 of the fully randomized “XXXY” sequence display a canonically feedback pattern of laminar activation, beginning in L1 and L6 before converging up through L4 to L2/3, followed by a feedforward pattern to L1 and L5/6 from L2/3. CSD was calculated over the average of n=77 complete oscillatory cycles out of 541 trials. Oscillatory LFP trace shown below for illustrative purposes only.

Unpredictable repetitions induce ɣ and θ oscillations.

The top row shows the two conditions contrasted to examine global oddball detection; the middle row shows the laminar signatures of the contrast results broken out by frequency band; and the bottom row shows the laminar CSD responses characteristic of each frequency band in these conditions. Solid contour lines enclose the area of a synchronization/power increase, while dotted contour lines enclose the area of a desynchronization/power decrease. (a) Spectral responses to P4 in an unpredicted (20% probability, n=2041 trials) repetition “xxxX” show a fast increase (2-4 dB) in ⍺/β (2-4 dB) and increase (<2 dB) in high-ɣ power, followed by a return to baseline in the ⍺/β band, a <2 dB increase in θ synchronization, and long-lasting <2 dB increases in low-ɣ (40-55 Hz) and high-ɣ (80-90 Hz) synchronization. (b) Spectral responses to P4 in a fully predictable (n=1355 trials) repetition sequence “xxxx” show a fast increase (4-6 dB) in ⍺/β power and smaller (<2 dB) ɣ power increase over baseline; both bands of oscillations continued throughout the first 200ms. (c) Global oddball detection, via the ⍺ = 0. 01 statistical contrast between “xxxX” and “xxxX”, induces a 0.2 dB decrease in ⍺/β power after 200ms, alongside 0.4-0.5 dB increases in θ and ɣ power. (d) ɣ power increases by 0.2-0.4 dB in L5/6, L4, and L2/3 immediately upon onset of the global oddball, persisting and even strengthening to 0.6 dB by the end of stimulus presentation, while ɣ power decreased by 0.4-0.6 in L1. (e) ⍺/β power increase begins (0.5-1 dB) in L5/6 <100ms after global oddball onset, while ⍺/β power decreases (0.5-0.75 dB) in L1. The increase in infragranular layers spreads out to L4 and then to L2/3 while weakening to 0.25-0.5 dB. (f) θ oscillations begin in L5/6 with a 0.5dB increase just after global oddball onset before spreading and strengthening (1-1.5 dB) throughout the laminar column. (g) ɣ oscillations induced by global oddballs show a canonically feedforward laminar activation pattern, starting in L4 at phase 0 and spreading out towards L2/3 and L5/6 as the phase cycle continues. CSD was calculated over the average of n=496 complete oscillatory windows over 2041 trials. (h) ⍺/β show a weakly feedback-driven laminar activation pattern, starting in L1 before continuing down to L2/3 and L4, with later contributions coming from L5/6. CSD was calculated over the average of n=300 complete oscillatory windows over 1355 trials. (i) θ oscillations show a feedforward laminar activation pattern under global oddballs, with activation beginning in L5 and L4 before spreading to L2/3 as it strengthens with the phase cycle. CSD was calculated over the average of n=335 complete oscillatory windows over 2041 trials.

Release from stimulus-specific adaptation induces a “push-pull” between

⍺/β- and ɣ oscillations. The top row shows the two conditions contrasted to examine release from stimulus-specific adaptation; the middle row shows the laminar signatures of the contrast results broken out by frequency band; and the bottom row shows the laminar CSD responses characteristic of each frequency band in these conditions. Solid contour lines enclose the area of a synchronization/power increase, while dotted contour lines enclose the area of a desynchronization/power decrease. (a) Spectral responses to P4 in a relatively predictable (80% probability) “xxxY” sequence conform to stereotyped “oddball” responses, with stimulus onset leading to an increase (2.5-5 dB) in ⍺/β power (followed by a return to baseline and then another, more sustained increase), accompanied by a fast-onset and then long-lasting increase (2.5-5 dB) in ɣ (40-55 Hz) and high-ɣ (80-90 Hz) synchronization. (b) Spectral responses to P4 in a fully predictable “yyyy” sequence show a fast increase (2-4 dB) in ⍺/β power after stimulus onset, followed by a smaller (<2 dB) increases in ɣ (40-55 Hz) and high-ɣ (80-90 Hz) synchronization. (c) Mostly-predictable release from adaptation, via the ⍺ = 0. 01 statistical contrast between “xxxY” and “yyyy” sequences, induces a 1-2 dB decrease in ⍺/β power at stimulus onset, followed by a 1-2 dB ɣ power increase and a roughly 1 dB increase in θ and ⍺/β power. (d) Release from stimulus-specific adaptation induces a fast, broad loss (<1 dB) of ɣ power, followed after 150ms by a 1 dB ɣ power increase from baseline that begins in L5/6 and spreads through L4 to L2/3. (e) Release from stimulus-specific adaptation induces a fast, broad loss (1-2 dB) of ⍺/β power, followed by a 1-2 dB increase from baseline in ⍺/β power that starts in L5/6 and spreads into L4 and L2/3 while never arriving to L1. (f) Release from stimulus-specific adaptation induces a fast, broad loss (1.5-3 dB) of θ synchronization that lasts longer in L2/3 than elsewhere, while after about 100ms a 1 dB increase in θ synchronization begins in L5/6 and spreads into L4 and L2/3 without ever arriving to L1. (g) ɣ oscillations in P4 of a mostly predictable “xxxY” sequence shows a canonical feedforward pattern of laminar activation, beginning in L4 and L5 at the end of one phase cycle and spreading out into L2/3 and L6 through the next cycle until the ɣ trough. CSD was calculated over the average of n=1910 complete oscillatory windows over 8462 trials. (h) ⍺/β oscillations in P4 of the fully predictable “yyyy” sequence show a feedback pattern of laminar activation, with current sinks spreading from L5/6 and L1 into L5, L4 and L2/3 before the end of the oscillation carries them into L1 and L6. CSD was calculated over the average of n=319 complete oscillatory windows over 1352 trials. (i) θ oscillations in P4 of the mostly predictable “xxxY” sequence show a mixed feedback and then feedforward pattern of laminar activity, with current sinks starting in L1 and L6 before activity in L2/3 at the θ trough spreads downward through L4 and L5 into L6 towards the cycle’s end. CSD was calculated over the average of n=1266 complete oscillatory windows over 8462 trials.

Early areas show greater, more significant differences in oscillatory ɣ power across deviance detection and global oddballs.

The bar-plot above shows the time-average of ɣ power increases/decreases in the statistically significant clusters (⍺ = 0. 01) shown in Figures 2(c), 3(c), and 4(c) above bars show the mean (over channels) increase/decrease for each contrast while the error marks show two standard errors of the mean (over channels). Deviance detection induced a significant increase in oscillatory ɣ power of roughly 1.5 dB in V1, 1.1 dB in LM, 1.3 dB in RL, 0.3 dB in AL, and insignificant differences in PM and AM. Global oddballs induced smaller, but statistically significant (for channelwise statistics) increases in ɣ-band power of 0.2 dB in V1, 0.1 dB in LM, and -0.5 dB in RL, with no significant contrasts in AL, PM, and AM. Release from stimulus-specific adaptation generated significant increases in ɣ-band power across area, averaging to 0.4 dB in V1, 1.2 dB in LM, 0.3 dB in RL, 1.2 dB in AL, 0.4 dB in PM, and 0.9 dB in AM.

Aperiodic components of spectral responses lack significant cross-condition differences.

Fitting parametric models using FOOOF (93)to the trial-averaged, within-subject (N=14) spectral responses to oddballs in areas V1, LM, RL, AL, PM, and AM separated out the aperiodic component (power shown in mV2/Hz in log-scale) from the broader spectral response; the six plots here each show this aperiodic component for the oddball (blue) and control (yellow) stimuli/sequences across the three contrasts, including error bands of two standard errors of the mean (SEMs). Across all three contrasts, the error bands for the oddball and control stimuli overlap closely enough to lack statistical significance. The blue, vertical dotted lines show the boundaries between θ (2-10 Hz), ⍺/β (10-30 Hz), and ɣ (30-90 Hz) frequency bands.

Oscillatory spectra show significant differences in power of ⍺/β and ɣ peaks in deviance detection and stimulus-specific adaptation contrasts across areas, while global oddball contrasts show limited significant differences.

After removing the aperiodic components shown in Supplementary Figure 1, the oscillatory peaks are shown here in log-space (decibels) across areas (V1, LM, RL, AL, PM, and AM) and contrasts (DD, GO, and SSA), with error-bands of two standard errors of the mean (SEMs) shown surrounding each average spectrum. DD and SSA consistently show significant (>2 SEM) differences in the heights (power above aperiodic baseline) of the ɣ (30-90 Hz) oscillatory peaks in V1, LM, RL, AL, and AM, with DD showing such a significant peak in PM as well. DD shows significant difference in heights of the ⍺/β (10-30 Hz) peaks, with the control condition showing greater ⍺/β power than the oddball condition, in V1, LM, RL, AL, PM, and AM; in LM and AL the SSA contrast shows higher ⍺/β power in the oddball condition rather than the control condition. In V1, LM, RL, AL, and PM the GO contrast shows higher ⍺/β power in the control condition. In DD, the oddball’s ɣ-band peak frequency significantly (in a permutation-based t-test with 1000 permutations) shifted higher than that of the control condition V1 (35-38 Hz, ⍺ < 0. 05) and LM (35-41 Hz, ⍺ < 0. 01). In SSA, the “oddball” (having probability 𝑃 = 80%) had a ɣ-band peak frequency shifted significantly (in a permutation-based t-test with 1000 permutations) to a lower frequency in LM (37-43 Hz, ⍺ < 0. 05), RL (37-45 Hz, ⍺ < 0. 01), AL (36-44 Hz, ⍺ < 0. 05), and AM (38-47 Hz, ⍺ < 0. 05). Statistics here were calculated within conditions and across subjects, averaging over channels and trials.

Aperiodic components of spectral responses in the first 150ms after oddball onset display significant cross-condition differences.

Fitting parametric models using FOOOF(93) to the trial-averaged, within-subject (N=14) first 150 ms of the spectral responses to oddballs in areas V1, LM, RL, AL, PM, and AM separated out the aperiodic component (power shown in mV2/Hz in log-scale) from the broader spectral response; the six plots here each show this aperiodic component for the oddball (blue) and control (yellow) stimuli/sequences across the three contrasts, including error bands of two standard errors of the mean (SEMs). Contrasts and areas marked with an asterisk (DD-LM, GO-AL, GO-PM, SSA-AM) displayed at least one cross-condition difference in the two parameters of the aperiodic component with significance level ⍺ < 5%, while those marked with two asterisks (GO-V1, SSA-V1, SSA-PM) displayed at least one cross-condition difference with significance level ⍺ < 1%; significance was measured by a permutation-based t-test on the two independent coordinates with 1000 permutations each. The blue, vertical dotted lines show the boundaries between θ (2-10 Hz), ⍺/β (10-30 Hz), and ɣ (30-90 Hz) frequency bands.

Deviance detection does not significantly change pupil size.

Taking the z-scores of pupil sizes across P4 of the “XXXY” random control sequences and “xxxy” habituation sequences did not show significant differences. Solid lines show the mean z-score of pupil size across trials, while error bands show two trialwise standard deviations of the mean (trialwise standard deviation over square-root of number of trials). The dotted line shows the mean pupil size overall with a z-score of zero.