WIP: What happens to intermediate representations as more layers are left at their random initialization? Here: the similarity between a convnet trained normally (all layers trained) and networks where only layers after layer n are trained (early layers untrained). 
Notice that the upper right quadrant remains roughly the same, even as the number of trained layers shrinks. Largest changes in off-diagonal similarity seems to occur below layer cnn7. This boundary may reflect bottom-up vs top-town information propagation?
Curious how this picture might look different using something like SVCCA. cc: @arimorcos @skornblith @maithra_raghu
Interesting that in my case (acoustic triphone recognition: 9000 classes), accuracy decreases gradually as more layers are left untrained, whereas accuracy drops quickly to 0% in @jasonyo et al (2014) after 3 random layers (in their case, visual object recognition).
I would love to see the similarity matrices for the visual networks with random layers. hmu if anyone wants to collaborate because this is just a fun side project and I need to graduate. 😅
