Spectral Networks and Locally Connected Networks on Graphs

Convolutional Neural Networks are extremely efficient architectures in imageand audio recognition tasks, thanks to their ability to exploit the localtranslational invariance of signal classes over their domain. In this paper weconsider possible generalizations of CNNs to signals defined on more generaldomains without the action of a translation group. In particular, we proposetwo constructions, one based upon a hierarchical clustering of the domain, andanother based on the spectrum of the graph Laplacian. We show throughexperiments that for low-dimensional graphs it is possible to learnconvolutional layers with a number of parameters independent of the input size,resulting in efficient deep architectures.

Further reading#

• Access Paper in arXiv.org