Explaining Deep Classification of Time-Series Data with Learned\n Prototypes

Abstract

The emergence of deep learning networks raises a need for explainable AI so\nthat users and domain experts can be confident applying them to high-risk\ndecisions. In this paper, we leverage data from the latent space induced by\ndeep learning models to learn stereotypical representations or "prototypes"\nduring training to elucidate the algorithmic decision-making process. We study\nhow leveraging prototypes effect classification decisions of two dimensional\ntime-series data in a few different settings: (1) electrocardiogram (ECG)\nwaveforms to detect clinical bradycardia, a slowing of heart rate, in preterm\ninfants, (2) respiration waveforms to detect apnea of prematurity, and (3)\naudio waveforms to classify spoken digits. We improve upon existing models by\noptimizing for increased prototype diversity and robustness, visualize how\nthese prototypes in the latent space are used by the model to distinguish\nclasses, and show that prototypes are capable of learning features on two\ndimensional time-series data to produce explainable insights during\nclassification tasks. We show that the prototypes are capable of learning\nreal-world features - bradycardia in ECG, apnea in respiration, and\narticulation in speech - as well as features within sub-classes. Our novel work\nleverages learned prototypical framework on two dimensional time-series data to\nproduce explainable insights during classification tasks.\n

Ähnliche Arbeiten