In classification tasks, the set of training examples for each class can be viewed as a limited sampling from an ideal infinite manifold of all sensible representants of this class. A layered artificial neural network model trained for such a task can then be interpreted as a stack of continuous transformations which gradually mold these complex manifolds from the original input space to simpler dissimilar internal representations on successive hidden layers – the so-called manifold disentaglement hypothesis. This, in turn, enables the final classification to be made in a linear fashion. We propose to assess the extent of this separation effect by introducing a class of measures based on the embedding complexity of the internal representations, with evaluation of the KL-divergence of t-distributed stochastic neighbour embedding (t-SNE) appearing as the most suitable method. Finally, we demonstrate the validity of the disentanglement hypothesis by measuring embedding complexity, classification accuracy and their relation on a sample of image classification datasets.
CITATION STYLE
Kuzma, T., & Farkaš, I. (2019). Embedding Complexity of Learned Representations in Neural Networks. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11728 LNCS, pp. 518–528). Springer Verlag. https://doi.org/10.1007/978-3-030-30484-3_42
Mendeley helps you to discover research relevant for your work.