Fig. 8
From: Creating interpretable deep learning models to identify species using environmental DNA sequences
The three closest test subsequences from any class to a given prototype sequence. X represents a base mismatch between the visualized prototype and the raw test sequence. As shown in the figure, at a latent weight of 0, each prototype matches perfectly with the closest test subsequences, but those test subsequences often belong to incorrect classes. On the other hand, at a higher latent weight, there are more base-pair mismatches between a prototype and its most similar test subsequences, but those test subsequences are more likely from the same class as the corresponding prototype. This also explains why higher latent weights produce higher accuracies, since they learn prototypes that better capture the most discriminative features of their assigned classes.
