04:59 < rcurtin> partobs-mdp: sorry about earlier, I didn't see your messages before you left; feel free to ask any off-topic questions at any time of course :)

05:02 < partobs-mdp> rcurtin: (Warning: a completely off-beat question) Could you advise some learning materials on assembly language? (preferably videos) I tried to self-study it several times, but failed every time.

05:02 < partobs-mdp> rcurtin: I know, it must be googleable, but the previous results show that something goes wrong in my case :(

05:06 < rcurtin> in my view it's easiest to start with a simple language

05:06 < rcurtin> or, rather, a simple architecture with few(ish) instructions

05:07 < rcurtin> when I learned assembly, it was a 'bottom-up' learning approach, where basically we started with "what are logic functions" and ended up with "how does a (primitive) processor work?" and from there, the assembly made sense

05:07 < rcurtin> I am sure there are better books, but the one I used was called ... hang on, let me find it, should be on my shelf here...

05:08 < rcurtin> "Logic and Computer Design Fundamentals", by M. Morris Mano and Charles R. Kime

05:08 < rcurtin> I'd loan you my copy but it might be too expensive to reasonably ship :)

05:08 < partobs-mdp> rcurtin: *** amazoning ***

05:09 < rcurtin> this was the first textbook I ever read like a novel... couldn't put it down

05:09 < partobs-mdp> rcurtin: Yes, it would be prohibitively expensive to deliver it from US to Moscow and from Moscow to Astrakhan :)

05:09 < rcurtin> it was my first semester of undergrad work

05:10 < rcurtin> but I think the book itself is a pretty typical introduction-to-computer-logic book, so there are surely many other books like it

05:13 < partobs-mdp> rcurtin: Well, the main thing is not to get the best possible book but to get the *good* book :)

05:14 < partobs-mdp> As far as I understand, the first six chapters discuss the computer design and all the rest is about the assembly?

05:15 < rcurtin> not sure if the new edition is the same; I had the third edition (but I guess that was 2005, there is probably a new edition)

05:15 < partobs-mdp> Looking at the preview, it looks really readable and able to fill my self-study blind spots.

05:15 < partobs-mdp> So a big thank you to rcurtin ^_^

05:15 < rcurtin> at least the edition I have, it's roughly the first 10 chapters that introduce how the processor works fully, and then after that they start talking about instructions, what they do, etc.

05:16 < rcurtin> and I think, once you got there, you could revisit learning an actual assembly language and go "oh! I get it now!"

05:16 < rcurtin> that's what happened to me---

05:16 < rcurtin> when I was in high school, I always wanted to write games for my TI-86 graphing calculator... but if you wanted to write good games you had to write assembly

05:16 < rcurtin> so I had my parents print off the entire TI-86 assembly programming manual, which was like 400 pages... and I didn't understand _any_ of it

05:16 < rcurtin> so I was stuck writing TI-BASIC to make games :)

05:17 < rcurtin> after a couple semesters of undergraduate study (including this book I'm referencing here) I revisited it and went oh! I get it! now I can do this!

05:17 < rcurtin> (but the interest was gone, I didn't work with a TI-86 regularly anymore, so I didn't ever write any assembly for it :))

05:18 < partobs-mdp> rcurtin: Wow, you had a really interesting high school + undergrad :) My high school was much less colorful

05:18 < partobs-mdp> rcurtin: You mean, the assembly begins to make a lot of sense iff you understand the processor design?

05:51 < rcurtin> partobs-mdp: yeah, that's what I meant, sorry for the slow response

05:51 < rcurtin> I don't think that my time with a TI-86 was 'really interesting', mostly I played with the TI-86 to avoid falling asleep (I wasn't usually successful) :)

05:51 < rcurtin> anyway thinking of sleep it is quite late here, so I will go to bed now; talk to you later

08:38 < partobs-mdp> rcurtin: Well, your time with a TI-86 should have been more interesting than my time with competitive programming, because that's virtually the only thing that I did in my last 2-3 years in high school :)

08:38 < partobs-mdp> rcurtin: Oh, and good night, by the way ^_^

10:51 < kris1> lozhnikov did you have a chance to look at the ssRBM PR yet.

11:04 < lozhnikov> kris1: No, I didn't. I'll point out the issue with NANs at github as soon as I find that

15:08 < ironstark> zoq: Thanks a lot for the help :)

19:06 < kris1> zoq: any particular reason that we have limited the output of binary cross entropy to between -5 and 5

19:18 < zoq> kris1: No, I think Konstantin did this just to test clipping.

19:18 < kris1> ok…..because i was using his implmenetation and the error was always between -5 and 5 so got confused

19:20 < zoq> kris1: https://github.com/mlpack/mlpack/pull/1070 Konstantin opened a new PR that includes the layer.

19:29 < kris1> ahhh….thanks. Just one more question why do outputlayer like crossEntropy layer require delta variable. I think the delta variable is used to backpropogate the error down to the layer but the delta of output layer is never used if i am not mistaken

19:38 < zoq> kris1: It's just how the network is implemented the error is stored in the layer which backpropagates the error,

19:45 < kris1> Ahhh okay one more thing so the error is backpropgated only till network.front + 1 i guess not till network.front. If i want to use the delta from the disriminator i think i would have to pass it through network.front and then pass the delta to the generator network.

19:47 < kris1> or should i just pass the delta/ error from the network[1] to generator.network.back() using the Backward function of generator.network.back()

19:48 < zoq> kris1: Yes, that would be the way to go, the reason for network.front + 1 is that there is no other layer that would need the delta to do the backward pass

19:51 < kris1> Sorry you mean to do the second option right.

19:53 < zoq> yes, right