Figure 5.

Abstract Factorized Codes Play Out in Synchrony with Replay

In addition to the decoding models trained to detect representations of individual objects (stim code), we also trained models to detect representations of “position within sequence” (pos code) and “which sequence” (seq code). See Method Details for details.

(A) During the applied learning phase of study 2, representations of seq, stim, and pos codes for the presented stimulus peaked at distinct times relative to stimulus onset.

(B and C) During the second resting period of study 2, spontaneous reactivation of both pos and seq codes consistently preceded spontaneous reactivation of the corresponding stim code, with ∼50-ms lag.

(D) As a validation, we also directly measured the relative timing of pos and seq activation and found a peak at 0 ms of lag; ^∗∗∗p < 0.001.

(E) Finally, we examined the temporal relationship between replay of stim codes (as shown in Figure 4B) and replay of pos codes (see Method Details for details). We found that pos code replay preceded stim code replay events by 50 ms.

(F) Summary of results. During each replay event, stimulus representations (green) were preceded 50 ms by abstract sequence (red) and position (blue) representations.

See also Figures S4 and S5.