Unlike classical associative memory models such as Hopfield networks, we are not limited in the expressivity of our energy model, and make use of the deep architectures with fully-connected, convolutional and recurrent layers.
Prikaži ovu nit
Our new work on memory uses a neural network's weights as fast and compressive associative storage. Reading from the memory is performed by approximate minimisation of the energy modelled by the network. https://arxiv.org/abs/1910.02720 pic.twitter.com/6uPAD9asgv
-
-
-
For this to work, stored patterns must be local minima of the energy. We use recent advances in gradient-based meta-learning to write into the memory such that this requirement approximately holds.pic.twitter.com/ASrOCZsHxd
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
-
-
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
You guys are too good and too fast. At this point I have lost the capacity to be shocked by your announcements.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Is the idea here to use this as a replacement for attention mechanisms?
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
#AI applying historical empirical to pattern heuristics perfecting the pattern match engine prediction matrix curating probailitic and/or persistent distortions. -
I don’t think I quite got your point
@DrSanjayPSahoo
Kraj razgovora
Novi razgovor -
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.