Can you always find a unique solution? That would be very surprising to me.
-
-
-
No there are cases where some weights or even neurons can not be identified. And a whole set of measure zero where horrible things happen.
Kraj razgovora
Novi razgovor -
-
-
This visualization is beautiful but there's something that seems a bit obvious about the fact that training a NN on the inputs and outputs of another NN would yield similar weights and biases. It's like aliens being surprised various species developed similar eyes independently.
-
Doesn't involve training a new NN actually - we calculate the unknown parameters from the boundaries between linear regions. I doubt that gradient descent would perform well here, even if one knew the architecture already.
Kraj razgovora
Novi razgovor -
-
-
That's pretty crazy, and suggests security issues as well
- Još 1 odgovor
Novi razgovor -
-
-
@gregeganSF Am I trippin? Is this a shape similar to the one you shared today?? -
The 2nd graph I posted, the Desargues graph, is a: https://en.wikipedia.org/wiki/Bipartite_graph … i.e. it contains two sets of vertices that have edges between the sets, but not within them. Most neural networks with two layers would, similarly, have edges between, but not within, the two layers.
- Još 1 odgovor
Novi razgovor -
-
-
Wow! I am impressed.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
