Yet, looking at the groups of networks with the same numbers of layers (see Figure 2 c) revealed a clear trend: For shallower networks, there is a positive correlation between accuracy and correlation with humans, meaning those exemplars that correlate well with humans are also the ones that perform better within the range of possible networks.
← all excerpts
Gloss perception: Searching for a deep neural network that behaves like humans.
1
—
—