On 12/06/2016 06:27 AM, Gael Varoquaux wrote:
On Sat, Dec 03, 2016 at 10:35:03PM +0000, federico vaggi wrote:
As long as the feature ordering has a meaningful spatial component (as is
almost always the case when you are dealing with raw pixels as features) CNNs
will almost always be better.
There is another important aspect: CNN are meant to work on data that is
translation invariant. Not all images are, for instance not brain images,
because they have been realigned. That said, most images are.
They also work on images that are not globally translation invariant.
Clearly you are the expert in brain images but the point is more that
the same feature detector make sense in multiple locations.
So as long as the local statistics and patterns are similar, I'd expect
it to work.
_______________________________________________
scikit-learn mailing list
[email protected]
https://mail.python.org/mailman/listinfo/scikit-learn
