Clustering and Anomaly Detection: Deep or Shallow Learning?

Abstract

In numerous classification applications, deep learning has achieved accuracy levels that were unimaginable just a few years ago. However, within the two primary branches of unsupervised learning —clustering and anomaly detection— the advantages of deep learning over traditional options are commonly uncertain or even non-existent. Why is this the case? Is unsupervised deep learning inadvisable when compared to shallow learning? Are they alternative or complementary approaches? In this talk, we aim to address these questions by delving into the characteristics and requirements of unsupervised learning, contrasting them with the limitations of deep learning and its current capabilities for these types of tasks.

Whether for describing, modeling, or predicting the behavior of any phenomenon captured in the form of data, unsupervised learning (whether as a central component or as a support) is indispensable in data analysis. It is a key element in the next AI, given the necessity for adaptive systems capable of facing novelty and the unknown.