Key problem for ML: train and test data share many superficial regularities with each other that they don't with the real world, so you can overfit the test data even without looking at it - and that's what deep learning does: https://arxiv.org/pdf/1711.11561.pdf …
-
This is inherent in statistical learning. The best we can hope for is to detect when new test data points violate these shared regularities. The power of fine tuning shows that the deep nets learn deeper regularities as well
But humans don't have this problem, so we must be doing something more.
-
-
*REALLY*? I never thought I did anything more than a multi-layer feedforward neural network! Back to the drawing board for me I guess ;-)
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
Humans don't stick with obsolete rules of thumb when the context changes around them?
-
Some humans do!
Keskustelun loppu
Uusi keskustelu -
-
-
Or maybe humans do have this problem, but our "training data" is sufficiently diverse and sufficiently resembles "real world data"
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
Yes, and we certainly don’t learn single tasks in isolation but instead learn many things simultaneously. The standard ML protocol is not robust
-
It goes past not being robust - it is categorically different from how organisms seem to learn from their environments best we can tell.
- Näytä vastaukset
Uusi keskustelu -
-
-
The problem isn't on statistical inference method but if data is representative of the process you are modeling
-
Agreed. But obtaining data that exhibit the full range of variability of current and FUTURE process behavior is very difficult and expensive.
Keskustelun loppu
Uusi keskustelu -
-
-
It's obvious that the human brain does not optimise an objective function. The brain does not even learn complex patterns. It learns only small elementary patterns that can be stitched together to instantly see any complex object even objects that it has never seen before.
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
I could be wrong but I don’t think the neurons in my brain are doing SGD to minimise a cost function.
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
Lataaminen näyttää kestävän hetken.
Twitter saattaa olla ruuhkautunut tai ongelma on muuten hetkellinen. Yritä uudelleen tai käy Twitterin tilasivulla saadaksesi lisätietoja.