In the Neuropixels dataset, almost every visual area was substantially biased toward decoding T RE as I C than L C , with V1, LM and AL showing the strongest representations of the IC
Figure 3d legend; Results, Representation of IC inference across the mouse visual cortical hierarchy, paragraph 5
What the paper reports
- V1, 12 sessions: Wilcoxon signed-rank test, p = 0.0132
- LM, 8 sessions: Wilcoxon signed-rank test, p = 0.0156
- RL, 11 sessions: Wilcoxon signed-rank test, p = 0.0244
- AL, 10 sessions: Wilcoxon signed-rank test, p = 0.0039
- PM, 12 sessions: Wilcoxon signed-rank test, p = 0.0342
- AM, 11 sessions: Wilcoxon signed-rank test, p = 0.1748
Under reading (b), adopted, the first seed gives four areas that decode TRE as IC significantly more than as LC (V1 p = 0.0024, LM 0.0156, RL 0.0049, AL 0.0020; PM 0.077, AM 0.083), with LM, V1 and AL the three largest mean scores (0.346, 0.308, 0.285). The two further seeds give five significant areas each, V1, LM, RL, AL and PM (PM p = 0.0425, AM p = 0.083) and V1, LM, RL, AL and AM (AM p = 0.0186, PM p = 0.052), with LM, V1 and AL the three largest. Under reading (a), five areas are significant with all three seeds (V1, LM, RL, AL and AM; PM p = 0.092 with the first seed) and V1, LM and AL are the three largest. The criterion is met in five of six runs. V1, LM, RL and AL are significant and LM, V1 and AL the three largest in all six; only the fifth area changes, and PM p runs from 0.0425 to 0.092 and AM p from 0.0020 to 0.083. The statement names PM as significant and AM as not; that set appears only with reading (b) and seed +100000, and the criterion counts areas without naming them. Sessions per area 12, 8, 11, 10, 12, 11, as in the paper.
How the test differed from the paper’s method
- Assumption. The Methods do not say how the SVM hyperparameters were set; the paper's code chooses them with OptimizeHyperparameters. Reading (b), C chosen by an inner cross-validation, is adopted as the reading closer to the code. It searches C from 10^-3 to 1 with one-vs-one coding. On the first outer fold of each session of V1, PM and AM, C above 1 gave a higher best inner accuracy in 1 of 35 sessions (PM, by 0.001), from fits that mostly did not converge (PM 28 to 30 and AM 30 of 33 or 36 per C value), and the one-vs-all coding changed the best inner accuracy by -0.026 to 0.027, which could change the chosen coding in some sessions; its effect on the inference scores was not computed (decoding_checks_shin_2025.ipynb). Asked of the authors under findings (methods_unavailable).
- Assumption. The cross-validation partition is random, and reading (b) meets the criterion with two of three seeds (the first seed gives four significant areas). The outcome follows the majority of the seeds, a rule chosen after the first seed's result and not stated in the criterion. A rule requiring every seed would give partially reproduced; averaging the per-session scores over the seeds was not computed.
- The decoder is scikit-learn's libsvm SVM, not MATLAB's fitcecoc.