Séminaire IMAGE : Edouard Oyallon (ISIR), « Learning is boring: image classification with patches »

I will discuss several old and recent results suggesting that simple baselines are still of interest even in large-scale Deep Learning applications. By simple, I mean those baselines are often shallow(with only a few ad-hoc layers), involve limited learning, and potentially some low-level representations like a dictionary of patches or cascade of wavelets. However, surprisingly, those representations are competitive on complex datasets.