⊘BlockedFigure 4D; Results, Distinct abstraction in CA1 and PFC, paragraph 2· Methods underspecifiedverifier disagreed at first
In PFC, the cross-condition generalization performance (CCGP) for trajectory phase, task sequence and environment is high and above trial-label shuffles for every decoder.
We found that the CCGP for all three task variables showed high accuracy for PFC
Figure 4D; Results, Distinct abstraction in CA1 and PFC, paragraph 2
What the paper reports
- trial-label permutation test, p < 0.001 (n = 1,000 shuffles)
With 8-cm single-trial maps 10 of the 16 PFC decoders have p < 0.001 by the code's strict count, but the task- sequence decoder tested on the inbound pair at the late segment predicts one class for every test point (CCGP 0.500, every shuffle ties, p = 1 counting ties). Counting ties, 9 of 16 have p < 0.001; the phase decoders tested on IN_R, OUT_R and OUT_L (CCGP 0.500, p >= 0.54 counting ties), that sequence decoder and three environment phase decoders (0.560, 0.560 and 0.541; p = 0.004, 0.004 and 0.020) fail, so the claim would not reproduce at 8 cm. Above every one of 100 shuffles are 14 of 16 decoders with 4-cm maps and 15 of 16 with 2-cm maps, and 4 of 16 without gap filling at 8 and at 2 cm; in this run no width gives all 16. In the verifier's round-2 copy, embedded on another machine before numba's CPU target was fixed, all 16 were above every one of 200 shuffles at 2 cm, so at the narrow widths the result sits at the edge of the criterion. The single-trial maps' smoothing is not in the Methods and the session width is garbled; at 8 cm the claim fails clearly and at 2 cm it can pass, and the claims file's rule for the width (fig2d-pv-similarity) makes the claim methods_underspecified.
How the test differed from the paper’s method
- UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.
- Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).
- The task-sequence decoders follow CCGP_taskSeq.m, whose test pairs take the two clusters at different segments (20-30% for one trajectory, 70-80% for the other); taking them at the same segment is a variant and gives CCGP within 0.003 of these.
- Linear C-SVM through scikit-learn's libsvm (C = 1), balanced accuracy, 1,000 label shuffles (decoding_CCGP_linearSVM_simple.m).
- The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity).
- A decoder that predicts one class for every test point has balanced accuracy 0.5 and ties with every shuffle; the code's p (fraction of shuffles strictly above) gives it 0. Ties are counted here (p_ge), so such decoders are not significant; the strict p is shown too.