Digital implementations of deep feature extractors are intrinsically informative

Rapid information (energy) propagation in deep feature extractors is crucial to balance computational complexity versus expressiveness as a representation of the input. We prove an upper bound for the speed of energy propagation in a unified framework that covers different neural network models, both over Euclidean and non-Euclidean domains. Additional structural information about the signal domain can be used to explicitly determine or improve the rate of decay. To illustrate this, we show global exponential energy decay for a range of 1) feature extractors with discrete-domain input signals, and 2) convolutional neural networks (CNNs) via scattering over locally compact abelian (LCA) groups.

Energy Propagation in Scattering Convolution Networks Can Be Arbitrarily Slow

We analyze energy decay for deep convolutional neural networks employed as feature extractors, such as Mallat's wavelet scattering transform. For time-frequency scattering transforms based on Gabor filters, it has been established that energy decay is exponential, for arbitrary square-integrable input signals. Our main results allow to prove that this is wrong for wavelet scattering in arbitrary dimensions. In this setting, the energy decay of the scattering transform acting on a generic square-integrable signal turns out to be arbitrarily slow. The fact that this behavior holds for dense subsets of $L^2(\mathbb{R}^d)$ emphasizes that fast energy decay is generally not a stable property of signals. We complement these findings with positive results allowing to conclude fast (up to exponential) energy decay for generalized Sobolev spaces that are tailored to the frequency localization of the underlying filter bank. Both negative and positive results highlight that energy decay in scattering networks critically depends on the interplay of the respective frequency localizations of the signal on the one hand, and of the employed filters on the other.