Cross-validation failure: small sample sizes lead to large error bars

Gael will give a talk at the next Neurospin unsupervised decoding meeting  on 21/02/2017 – 9h45-11h00, room 2032

Abstract:  Recently, I have become convinced that cross-validation on a hundred or
less samples is not a reliable measure of predictive accuracy. In
addition, techniques generally used to estimate its error or test for
significant prediction are severely optimistic.

I would like to present you very simple evidence of this unreliability,
which it intrinsic to the sample sizes that we are working with. It is a
simple sampling-noise problem that cannot be alleviated without increasing
the number of samples.

I want to have a discussion about what this means for the field, and how
we should address this problem. I would like to invite critical thinking
about aspects of the practice that I might have overlooked and could make
it more robust.

I invite many people to come, so that we can convince ourselves of
whether or not there is a problem with the way we often work. Indeed, it
is troublesome for methods development as well as for neuroscience

Comments are closed