Topic: data manifolds

Two very successful solutions exist to the problem of vision: biological brains and convolutional neural networks (CNNs). Despite the inspiration of artificial architectures from their biological counterparts, the extent to which the two solutions are comparable remains unclear, with evidence pointing to both core similarities (e.g., hierarchical learning of feature detectors with increasingly complex tuning) and key differences (e.g., very different sensitivity to pixel-level noise). In our work, we present results that bridge these two worlds, showing that prominent information processing tr