Supplementary Figure 2: Cross-temporal decoding in the LPFC and FEF.

(a, b) The cross-temporal decoding performance, LP ₁₁ (50.5 ± 0.9%), is shown for classifiers trained and tested with data from Delay 1 (800 – 1300 ms after target onset, highlighted by the red bar in the bottom plot, which is the average performance in the square region labelled LP ₁₁ in Fig. 2a). When these classifiers were tested in Delay 2 (2100 – 2600 ms after target onset), the performance, LP ₁₂ , was 32.4 ± 1.0% (the magenta bar in the bottom plot). This was still above chance (which was 1 out 7 = 14.3%), but significantly lower than LP ₁₁ . From these results, it was clear that the distractor caused a significant drop in the performance of the classifiers trained in Delay 1 (shown in the middle box plot in (b), p < 0.001). However, it was not clear if there was an overall decrease in target information in Delay 2 after the presentation of the distractor, or if the target information was now encoded differently. To answer this question, we looked at classifiers trained and tested in the same 500 ms window in Delay 2. We found LP ₂₂ to be 47.9 ± 0.9% (shown in the red bar in the top plot), which was not significantly different from LP ₁₁ (the difference is shown in the left box plot in (b), p ~ 0.07). This showed that there was very little loss in information in Delay 2 after the presentation of the distractor. This meant that the reason for the drop in performance for the Delay 1 classifiers when tested in Delay 2, was a morphing in the population code in Delay 2 that encoded target information differently from Delay 1, and not due to an overall decrease in target information in Delay 2. Interestingly, if the classifiers trained in Delay 2 were tested on the data in Delay 1, the performance decreased significantly again (LP ₂₁ = 37.2 ± 0.1%, shown in the magenta bar in the top plot). The change in performance is again plotted in (b) (right box plot, p < 0.001). This meant that the morphed code in Delay 2 incorporated very little of the code in Delay 1, resulting in significantly poorer performance when tested on the responses in Delay 1, suggesting a fairly large transformation occurred. The error bars in (b) indicate the 95th percentile range (two sided) obtained through 1000 measurements of the performance. (c) The shift in the cluster centers from Delay 1 to Delay 2 (i.e. the blue to the red dots in Fig. 3a & b) for each cluster is plotted here. All the shifts found in the Delay 1 space and Delay 2 space were significantly larger than chance (the 97.5^th percentile of the shifts obtained by chance is denoted by the dashed line, p < 0.001 for all locations, from left to right. (d – e) Same plots as (a, b), but for the FEF. The performance of classifiers trained in Delay 1, went from FP ₁₁ = 39.0 ± 0.9% when tested in Delay 1, to FP ₁₂ = 28.8 ± 0.7% when tested in Delay 2 (the difference is shown in the middle box plot in (e), p < 0.001). This appeared to be similar to what we observed in the LPFC. However, when we trained and tested classifiers in Delay 2, the performance was found to be FP ₂₂ = 31.7 ± 0.7% (the difference is shown in the left box plot in (e), p < 0.001). These results suggest that there was an overall loss in information in Delay 2, consistent with a degraded code, rather than a change in the encoding of the information. Surprisingly, the classifiers trained in Delay 2 when tested in Delay 1 produced a performance of FP ₂₁ = 30.6 ± 0.1% in Delay 1, which was not significantly different from the performance found in Delay 2 (the difference is shown in the right box plot in (e), p ~ 0.06). This was quite different from what happened in the LPFC, where we observed a significant drop from LP ₂₂ to LP ₂₁ . This appears to be an interesting asymmetry in the performance that warrants further study. (f) The shift in the cluster centers for each cluster is plotted here. All the shifts were not significantly larger than chance (p ~ 0.75, ~ 0.98, ~ 0.51, ~ 0.81, ~ 0.94, ~ 0.30, ~ 0.97, and ~ 0.67 for Delay 1; ~ 0.99, ~ 0.12, ~ 0.48, ~ 0.34, ~ 0.47, and ~ 0.16 for Delay 2). n.s. - not significant. ** - indicates that the 2.5^th and the 97.5^th percentile range did not overlap with 0 (in b and e) or with the 95th percentile range obtained from a chance distribution. The horizontal red line in the box plots represents the median of the distribution.