09:28 < clicksaswat> hi guys! i submitted a proposal for gsoc under the title "Research and Implementation of Augmented RNNs", but due to timing constraints, never had a chance to ask for feedbacks, if you guys have time for it now, any comments for the proposal would be really appreciated.

12:20 < zoq> clicksaswat: I can comment on the proposal, if you like.

13:54 < Trion> zoq: commented the error in https://gist.github.com/zoq/c3e4f90740399c6d42ac9e6f66a2c53c

14:17 < clicksaswat> zoq: hey zoq! that would be really great. I would like to hear especially about the part where I've mentioned extending RNN architectures to support for external memory. Is it feasible? Also, if you have comments/doubts about any other parts of the proposal, feel free to mention it.

21:06 < rcurtin> no word from the team that is supposed to be installing masterblaster...

21:06 < rcurtin> don't know if they will have it done before they leave tonight

21:06 < rcurtin> if they don't that means downtime until probably Monday or Tuesday

21:07 < rcurtin> but... at least it is better than the year it took to get it online...

22:21 < sumedhghaisas> rcurtin: Hey Ryan... did you get a chance to look at that softmax regression problem?

22:21 < sumedhghaisas> :)

22:27 < rcurtin> it's on my queue

22:27 < rcurtin> actually pretty close to the top finally :)

22:28 < sumedhghaisas> haha... cause I was thinking... I am right now implementing a neural net architecture nd suddenly realized in classification we always have a softmax layer on top

22:28 < sumedhghaisas> so I have this mathematical question

22:28 < sumedhghaisas> if we call our thing a reduced softmax...

22:29 < sumedhghaisas> softmax has many optimums... as we can subtract any vector from the weights

22:29 < sumedhghaisas> but reduced softmax has only 1 optimum

22:29 < sumedhghaisas> thus an optimization should be more stable in reduced softmax right?

22:30 < sumedhghaisas> I am not so sure about it but it feels right