Figure 11: Real-time inference and time-varying features.

a, Inference speeds as a function of video resolution for both DeepLabCut live and proSVD during streaming analyses for different downsampled video resolutions. Error bars denote standard deviation (N=10). b, The L2-norm of the derivative of the first feature identified by proSVD plotted as a function of time post-stimulation, averaged across all trials. The low-dimensional representation of the evoked motion shows some different temporal features for each stimulus conditions.