You may have noticed that our magazine covers have been venturing beyond the customary look of a technical publication. I hope this shift has been enjoyable for you to see. I confess that I’ve been enjoying it myself, particularly the process of selecting the images and shaping the final composition. It has been an engaging experience exploring the possibilities and connections between the ideas we wish to highlight and images drawn from the natural world.
In this instance, our cover image for the special issue “Mathematics of Deep Learning” features a canyon, which serves as a geometric metaphor for the mathematical landscape of deep learning. Among the many canyon images we considered, Blyde Canyon emerged as the favorite. Its shape nicely evokes the valleys and ridges of rugged loss surfaces, and it is less instantly recognizable than landmarks like the Grand Canyon. As an added bonus, its contours even appear reasonably differentiable—just enough to keep our optimization metaphors comfortably aligned with the practice of deep learning.
In the same way a canyon is sculpted by persistent local forces, a deep network takes its shape through such repeated updates, and the global vast pattern emerges from many small, systematic steps. The river flowing naturally through this beautiful landscape with gentle slopes follows the path of steepest descent, the workhorse of optimization in deep nets. Finally, a canyon represents an accumulated transformation with geological layers very much like deep networks where each layer encodes a different aspect from simple features to an increasingly abstract structure.
