Extended Data Fig. 7: Quantification of diversity and bipartite invariance indices from electrophysiological data.
From: Functional bipartite invariance in mouse primary visual cortex receptive fields
a, Normalized correlation coefficients (CCnorm) of the model fine-tuned from a pretrained core on two-photon recordings (denoted as ‘pretrained 2p core’) were higher than those from the model trained from scratch (denoted as ‘end-to-end’) (two-sided Wilcoxon signed-rank test, W = 10762, P = 2.1 × 10−24) with color representing units classified as either ‘single units’ or ‘multi-unit activity’ (denoted as ‘Neuropixels single’ and ‘Neuropixels multi-unit activity’ (Neuropixels MUA), respectively). b, Histogram of the normalized correlation coefficient (CCnorm) for neurons modeled from two-photon recording (denoted as ‘2p’, data were pooled over 33,714 neurons from 14 mice), and Neuropixels units modeled using electrophysiological data with pretrained 2p core (data were pooled over 364 spikes sorted units from six mice). Excessively noisy neurons (CCmax < 0.1) were excluded (0.2%, 7.6%, and 4.6% of 2p, Neuropixels single, Neuropixels MUA, respectively) and neurons with values outside of 0 and 1 were clipped (1.2%, 4.5%, and 6.1% for 2p, Neuropixels single, and Neuropixels MUA, respectively) for visualization. c, Neuropixels MUA (0.88±0.29) exhibited larger diversity indices than Neuropixels single units (0.55±0.31) and 2p neurons (0.51±0.27) (P = 9.3 × 10−16 and 1.7 × 10−37, two-sided Welch’s t-test with 230.3 and 122.6 d.f., respectively) while Neuropixels single units had similar diversity indices to 2p neurons (P = 0.21, two-sided Welch’s t-test with 158.4 d.f.). d, Neuropixels MUA showed larger bipartite invariance indices (BIIs, 0.74±0.08) than Neuropixels single units (0.62±0.10) and 2p neurons (0.65±0.08) (P = 7.5 × 10−21 and 1.3 × 10−22, two-sided Welch’s t-test with 237.8 and 129.5 d.f., respectively) while Neuropixels single units had marginally lower bipartite invariance indices than 2p neurons (P = 2.8 × 10−6, two-sided Welch’s t-test with 155.7 d.f.). e, Neuropixels single units (38.8±6.36°) had larger RF sizes than both Neuropixels MUA (35.6±4.99°) and 2p neurons (34.9±6.64°) (P = 3.3 × 10−5 and 1.7 × 10−10, two-sided Welch’s t-test with 235.5 and 161.3 d.f., respectively) while MUA had similar RF sizes to those of 2p neurons (P = 0.30, two-sided Welch’s t-test with 141.6 d.f.). f, Histogram of the common logarithm of the inter-spike-interval (ISI) violations for Neuropixels single and Neuropixels MUA. Neuropixels MUA exhibited high ISI violations (median 0.94), whereas Neuropixels single units displayed a smaller mean but broader distribution of ISI violations (median 0.09). g, Diversity indices were not correlated with the ISI violations for Neuropixels single units (Pearson r = − 0.03 with p = 0.76). h, Bipartite invariance indices were not correlated with the ISI violations for Neuropixels single units (Pearson r = − 0.05 with p = 0.55). g–h, Units with excessively large ISI violation (> 10) were excluded for visualization (two single units and one MUA). c–h, Neurons were selected to have oracle scores larger than 0.22 and model test correlations larger than 0.42. Two-photon data were pooled over 1,154 neurons from six mice; electrophysiological data pooled over 240 spike sorted units from six mice.
