Fig. 3: DeepHL analysis of repulsive odor learning in worms.

a The experimental setup (left) for monitoring the worm’s trajectory (right). b Example trajectories of worms colored by attentions of a discriminator layer. Segments of the trajectories corresponding to the run state of the worm are highlighted (in red). c Time series of the moving average of speed (black lines) associated with attention values (colored lines). The upper and lower graphs are obtained from the upper and lower trajectories shown in b, respectively. d Histograms showing the distributions of the moving variance of speed for each time slice within the highlighted trajectory segments. e Frequency analysis of the velocity of a preexposed or control worm. A 128-s-wide sliding window was shifted in 1-sample intervals and the amplitude of each frequency component was obtained from its fast Fourier transform (FFT). The upper and lower spectrograms were, respectively, obtained from the upper and lower trajectories shown in b. f Frequency analysis of the velocity of all the preexposed or control worms computed from entire trajectories. The histograms and box plot show the distributions of the dominant frequency of speed for each time slice. The dominant frequency is the one with the largest amplitude within each window. Significant difference in the dominant frequencies were observed by a generalized linear mixed model (GLMM) with Gaussian distributions (t = −6.60; d.f. = 322.8; p = 1.68 × 10⁻¹⁰, effect size(r²) = 0.232; **p < 0.01; see “Methods”). The p value is two sided. The box plot shows the 25–75% quartile, with embedded bar representing the median; lower whiskers show Q1 − 1.5 × IQR (Q1: 25% quartile; IQR: interquartile range); upper whiskers show Q3 + 1.5 × IQR (Q3: 75% quartile). Control: n = 76, 784, preexposed: n = 75, 750.