Fig 4. Fine-tuned APD outperforms both naive APD and MSE.

A) An example set of morph stimuli, interpolated between stimuli B (left) and C (right). Refer to [26] for a detailed description of stimuli generation. Briefly, the author linearly interpolated between low-dimensional representations of B and C, and reverted the interpolation vectors to spectrograms. B) Computed and behaviorally measured psychometric curves on example morph stimuli shown in A. APD (naive) and APD (tuned) are both calculated from APD feature vectors, with the former only pre-trained (naive) and the latter fine-tuned on animal behavior data (tuned). C) Pairwise error in Hill coefficient measurements between each distance metric and the ground truth. For each computed psychometric curve (MSE, naive APD, and tuned APD), we calculate the error between its computed Hill coefficient and the ground truth value under the same training conditions (morph stimuli, cohort, etc.). For the ground truth, we calculate its internal variability by measuring errors between all pairs of subject judgments under the same training conditions. Outliers are not plotted. D) Pairwise error in inflection point measurements between distance metrics and ground truths. Error calculation follows the same pattern mentioned in C. All computed psychometric curves yield measurements within the variability of ground truths. Outliers are not plotted.