Intelligence is in the human trainer who looks for the right number of layers and parameters to obtain the best compression that shows a semblance of generalization for other inputs
Gradient descent is just a convex optimizer. By definition, it can't solve hard problems. (Fitting the data means nothing; memorization is enough for that.) If GD is part of an AI system, the intelligence must be somewhere else.
-
-
-
Not enough.
- Näytä vastaukset
Uusi keskustelu -
-
-
The intelligence lies in the selection of a proper cost function.
-
But the current ones (GANs aside) are so simplistic it’s hard to take that seriously.
Keskustelun loppu
Uusi keskustelu -
-
-
You think it is literally impossible to build a turing complete system with multi dimensional 'convex optimizers'?
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
Every curve is convex in some space
-
Finding that space is what takes a lot of smarts.
- Näytä vastaukset
Uusi keskustelu -
-
-
What's your take on Neuro-Symbolic AI?
-
Symbolic AI when done properly is the way AGI will be achieved. Spatiotemporal invariant concepts that are bootstrapped as the algorithm experiences. Starting with edges and gradients upto the most complex of ideas. All grounded and contexted.
Keskustelun loppu
Uusi keskustelu -
-
-
GD is not just a convex optimizer.
-
Could you elaborate Dan? Gradient descent takes a local greedy decision. If the function is convex then it will find the local minimum, if the step is ok. But the fact that the step can make the algorithm overshoot is not finally a bug but a feature because it can escape it.
- Näytä vastaukset
Uusi keskustelu -
Lataaminen näyttää kestävän hetken.
Twitter saattaa olla ruuhkautunut tai ongelma on muuten hetkellinen. Yritä uudelleen tai käy Twitterin tilasivulla saadaksesi lisätietoja.