Gary Marcus

Invariance isn’t just over-rated, worse: enforcing invariance in a design can actually be harmful to general task performance. Avoid if possible. So I think I agree with you there, and raise you.

Come on guys. Invariance is an absolute must. There can be no intelligence without it. The brain does not use lots of training samples to achieve it. The brain uses timing as a glue. Transformations of an object do not break invariance because the brain can predict future states.

But invariance should be learned, not imposed as a prior. That is the point. Different signals have different invariances. Translation is just one of many...

why learned?

Because I don't think we can easily tell what the invariances are that we'd like to put in (particularly higher up in the sensory hierarchy). The system needs to be constructed in a way that would allow for it to pick them up. See ferret cortical rewiring experiment.

I would think that we need some invariances innately, others to be acquired. If you are talking about Sur’s rewiring experiments, I don’t see the relevance