I'm having trouble understanding why the success of pooling would be deemed unfortunate. Max-pooling or average-pooling are essentially learning something about a pair of contiguous features: while max-pooling saves only the most prominent/largest value, average pooling compresses information using the mean. Saying this is unfortunate from Hinton's standpoint amounts to saying that it is very unlikely to see any sort of pooling behavior at the level of neuronal populations. At a physiological and psychological level, what would pooling equate to?
I think the intuition behind max-pooling is that it says "something is in this local region", without overly specifically saying where it is. Intuitively, a human may detects an edge or intense bright light, but not care so much about precisely where it is in the field of vision. After many successive layers of max-pooling, however (if the pools are not over-lapping), even somewhat course information about locality is lost.
I believe Hinton objects to this gross loss of spatial information for two reasons:
1) Humans don't lose so much spatial information, and Hinton would like his models to ultimately capture a neurologically plausible computation.
2) It may not be necessary for object detection (Imagenet), but it would likely be important for more sophisticated tasks.
He also 'did not like' Support Vector Machines, back when they were the best method for image recognition. His reason was that SVMs were a 'dead end' - they were not a step on the path to human-level image recognition. His argument now seems pretty valid.
I think he is saying the same thing about Max Pooling. Just my guess.
The AMA highlights one of the deficiencies of Markdown. Prof Hinton has attempted to number many of his responses to the highest rated thread, but they're all coming up as 1. 1. 1. 1.
>I guess we should just train an RNN to output a caption so that it can tell us what it thinks is there. Then maybe the philosophers and cognitive scientists will stop telling us what our nets cannot do.
I wonder if he knew about the Stanford paper that demonstrates this? Or if he just guessed this would happen.
7 comments
[ 2.9 ms ] story [ 20.7 ms ] threadI believe Hinton objects to this gross loss of spatial information for two reasons: 1) Humans don't lose so much spatial information, and Hinton would like his models to ultimately capture a neurologically plausible computation. 2) It may not be necessary for object detection (Imagenet), but it would likely be important for more sophisticated tasks.
I think he is saying the same thing about Max Pooling. Just my guess.
I wonder if he knew about the Stanford paper that demonstrates this? Or if he just guessed this would happen.
http://cs.stanford.edu/people/karpathy/deepimagesent/
http://googleresearch.blogspot.ca/2014/11/a-picture-is-worth...
Or the Toronto one? http://deeplearning.cs.toronto.edu/i2t
Two institutions he is affiliated with.