Tea Talk 29 Oct Wed @13.30 AA3195 by Dong-Hyun Lee
Dear all, Dong-Hyun will update us on the progress in using target-prop this week at the tea talk. The talk will be held at the usual place AA3195 and usual time of 13.30. Hope many will attend the talk! - Cho === - Speaker: Dong-Hyun Lee - Title: Training supervised deep network and an auto-encoder through target propagation without back-prop - Abstract: I'll present how to train supervised network through target propagation. This new technique suggested by Yoshua can make it possible to train deep network without back propagation thanks to layer-wise target matching losses and target propagation through layers. Our results are comparable to back-propagation on MNIST. And we can train discrete networks directly which is impossible to train using ordinary back-propagation because this technique doesn't use any derivatives for propagating training signals. Using this, we can also train an auto-encoder without back-prop.
Afficher les réponses par date
Dear all, I'll present "Training supervised deep network through target propagation without back-prop" at AA3195 and 1:30pm. This is slides for tea talk today. Unfortunately, I decided not to present backprop-free auto-encoder in order not to spend too long time. Feel free to ask it in any time. Dong-Hyun. 2014-10-28 6:45 GMT-04:00 Kyung Hyun Cho <cho.k.hyun@gmail.com>:
Dear all,
Dong-Hyun will update us on the progress in using target-prop this week at the tea talk. The talk will be held at the usual place AA3195 and usual time of 13.30.
Hope many will attend the talk! - Cho
=== - Speaker: Dong-Hyun Lee - Title: Training supervised deep network and an auto-encoder through target propagation without back-prop - Abstract: I'll present how to train supervised network through target propagation. This new technique suggested by Yoshua can make it possible to train deep network without back propagation thanks to layer-wise target matching losses and target propagation through layers. Our results are comparable to back-propagation on MNIST. And we can train discrete networks directly which is impossible to train using ordinary back-propagation because this technique doesn't use any derivatives for propagating training signals. Using this, we can also train an auto-encoder without back-prop.
_______________________________________________ Lisa_labo mailing list Lisa_labo@iro.umontreal.ca https://webmail.iro.umontreal.ca/mailman/listinfo/lisa_labo
-- - Dong-Hyun
participants (2)
-
Dong-Hyun Lee -
Kyung Hyun Cho