Top Highlights

• From the generated movies, frames were manually selected for inclusion in the dataset with the goal of maximizing the representation of diverse morphologies and poses from within each movie. Each frame was then manually labeled and the labeled coordinates were used to generate target heatmaps (i.e., two-dimensional Gaussians) that served as the “gr...
• Datasets were randomly divided into training (75%) and validation (25%) splits. Image pixel intensities were shifted from the integral range 0–255 to the floating point range 0.0–1.0. During training, we made use of data augmen- tation (Shorten & Khoshgoftaar, 2019) by randomly performing vertical and horizontal flips, rotating up to 45 degrees clo...
• Training was performed using mean squared error (MSE) loss, batch size equal to 8 (touch-evoked response dataset) or 32 (visuomotor response dataset), and the Adam optimizer
• In the first approach, we adopted an hourglass-shaped architecture. In this architecture, as information flows through the successive layers of the network, it is first downsampled, then upsampled. The downsampling or “encoding” path helps to minimize the computational cost of the expensive convolutional layers and pack the infor- mation into a com...
• Head-stimulated slc1a2btk57/tk57 embryos swim significantly longer and farther than wild-type controls (two-way ANOVA, F(1,86) = 59.58, p < 0.001 and F(1,86) = 57.81, p < 0.001 for duration and distance, respectively

Domain