Fig. 5
From: Creating interpretable deep learning models to identify species using environmental DNA sequences
Test accuracies from ProtoPNet models trained at different latent weights. These models used prototype length 5, and were trained on noise level 1 and tested on noise level 0. This shows that slightly lowering the reliance on the convolved output actually increases accuracy. For latent weights lower than 0.5, the higher the reliance on the convolved input, the better the test accuracy. However, for latent weights above 0.5, the test accuracies were within 1% of each other. Choosing a lower latent weight makes decisions more interpretable, so we wanted to choose the lowest latent weight without sacrificing significant accuracy. We selected a latent weight of 0.7 based on the validation data, but the test results here show that the latent weight could have been lowered to as low as 0.5 without a significant penalty to accuracy. Note that the y-axis spans from 88 to 97% accuracy for purposes of comparison. Error bars are standard deviation for n = 5 runs.
