Shane Chu

Understand convolutional dictionary learning vs convolutional neural network in one picture

In his talk (at 21:57), Brendt Wholberg illustrates the distinction between Convolutional Dictionary Learning (CDL) and Convolutional Neural Networks (CNNs) with the following slide:

Slide from Brendt Wholberg's talk on Convolutional Sparse Representations for Imaging Inverse Problems at NC state ECE 2018.

In CDL, the filters are convolved with code vectors to represent the input signal, akin to a deconvolutional network that provides a translation-invariant sparse representation.

On the other hand, in CNNs, filters are convolved with the input signal (such as an image) to generate a feature representation.

References

• Matthew D and Krishnan Zeiler, Deconvolutional networks, 2010.
• Cristina and Wohlberg Garcia-Cardona, Convolutional dictionary learning: A comparative review and new algorithms, 2018.