Fig. 4: The stability of late-delay epoch representation.
From: Volatile working memory representations crystallize with practice
a, Schematic of LBM (left). A pump pulse is split into 30 light beads, each delayed (Δτ ≈ 7 ns, where Δτ represents the time delay between excitation foci) and focused at different sample depths, enabling full-volume sampling at the microscope’s frame rate. Top right, a 3D rendering of neuron locations across 30 planes. Bottom middle, the activity of 36,471 neurons during A-odour or B-odour trials, sorted by response and shown over 60 min. Bottom right, a 60 min raster of these neurons, highlighting ΔF/F activity for selected neurons. Scale bar, 500 μm. b, The accuracy of decoding the first odour and choice during the first-odour, late-delay and choice epochs using the activity of neurons at different depths. n = 4 mice. c, The performance of the mouse during the 10 days of calcium imaging. Mice (n = 4) were performing at expert level from the first day of imaging. d, Example field of view showing 1 out of the 30 planes with the same constellation of neurons across multiple imaging days. e, Decoding first odour and choice over 10 days with a daily-trained decoder, using up to 73,307 neurons across 30 planes and 450 μm depth (top) (n = 4 mice). Bottom, quantification of accuracy by epoch; the dashed lines indicate shuffled data decoding. f, Similar to e, with the decoder trained on day 10 and tested across days 1–10. g, The decoding accuracy for first odour during first-odour and late-delay epochs, and choice accuracy during the choice epoch. n = 4 mice. h, The average across-day decoding accuracies for early versus late days (n = 4 mice), as depicted in the off-diagonal terms of Fig. 4g, with significant improvements (first-odour and late-delay, P < 0.0001; choice, P = 0.0003). Statistical analysis was performed using unpaired t-tests. For b, c, e and f, data are mean ± s.e.m. Details of the statistical analyses are provided in the Methods.
