Demian Till comments on Do sparse autoencoders find “true features”?

[Demian Till]

My bad! Yes since that’s just one batch it does indeed come out as quadratic overall. I’ll have a think about more efficient methods