sinwave

Lol - the code example seems to imply that Zuckerburg's user id is 4.

Geoffrey Hinton gave results to this effect in a talk about "Dark Knowledge" [1]. Haven't seen any of these results published, though. I think he mentions something in the talk about NIPS rejecting the paper.

[1] - https://www.youtube.com/watch?v=EK61htlw8hY

Thought I'd chime in that for image compression (e.g in the JPEG2000 standard), the 2D discrete wavelet transform takes advantage of similar pixel intensities for neighboring pixels at various scales (i.e. "transformed magnitudes all nice and grouped for you"). The 2D-DWT is actually pretty cool under the hood. And, asymptotically, a bit faster than the FFT (DWT runs in O(N), and in 2D, O(width*height)).

May I please have your faster than nlogn sorting algorithm

This is actually some really great advice to hear upon graduating from college. Thank you.

This is really cool work. There is a lackluster theano feature which allows you to print a flowchart figure for the "computation graph" corresponding to the symbolic representation of your model.

@ajtulloch's library provides what I imagined the feature would be at first glance - a comprehensible, elegant graphical representation of your NN model. And on top of that, all in Haskell, with Haskell DSL for running torch - so cool.

Yeah, imagining him touring while writing his thesis has been a source of inspiration lately.

Am I wrong that the title seems to imply something negative about word vectors? But the article is super pumped about them!

There is a seemingly dead github repo called "R2D3" by hadley wickham. Wonder what came of that project?

https://github.com/hadley/r2d3

cf. Searle's Chinese Room argument!

I know! According to [1] the minimax algorithm does indeed originate here. And Hamming said in that talk [2][3] that Shannon didn't do anything killer after Info Theory...

[1] - http://webdocs.cs.ualberta.ca/~sutton/book/the-book.html [2] - http://www.cs.virginia.edu/~robins/YouAndYourResearch.html [3] - https://www.youtube.com/watch?v=a1zDuOPkMSw

Alarm radio wakes me up with jazz

{run,don't run}

make coffee/tidy up/drink coffee

shower

make breakfast

eat breakfast over coffee and {math textbook, video lecture}

commute on bike

This reminds me of an opinion that I have which I've been looking for an opportunity to flesh out and share. Despite all of the research and engineering that's gone into the user interface (and "experience") of smartphones, text messaging as an interface has one massive advantage: the perceived cost of sending a text is miniscule in comparison to other operations. I don't have units, but it's probably an order of magnitude lower in any reasonable ones.

Maybe I'm projecting onto "the general public" when I make this generalization, but performing operations on a smartphone (aside from call/text) are oft accompanied by the very real risk of squandering your time away. Especially when a browser is involved.

There is something elegant about the interface of a dumb phone, especially a flip phone: You pull it out of your pocket, whip it open, type your text, send it. And then crucially, you flip it closed and stop thinking about it. This is the key. You're using it when you're using it, and you're not when you're not.

Even grander generalization and possibly controversial opinion: I hope that consumer technology begins to cater more to those of us who wish to use technology in this way - as a tool that you pull out of your pocket and promptly put away upon achieving your ends.

Incidentally I'm reading Sipser's book for a philosophy course right now.

wait - are they actually building aircraft carriers??

Hilarious moment at 36:00. Sutton has just finished talking about how in real life, functions that learning agents wish to estimate will often be non-stationary because life is constantly changing.

An audience member complains that this is unfair (i.e. unfair to the learning algorithm), and somebody nearby the cameraman retorts "life is unfair".

This is a fact that even algorithms cannot avoid.

I suspect that the best way to tackle this problem is to treat "learning to walk" as a reinforcement learning problem. This way you can evaluate actions within each episode rather than waiting until the end of each trial to adjust parameters encoding the walking strategy.

As karpathy suggests below (and he's certainly much more qualified than me), an evolutionary method such as GA's, while apparently fairly effective, could well be wasting valuable information learned in real time through interaction.

The "Machine Learning Summer School" lecture series is happening now at UT Austin - I made the call to catch up with family over my winter break rather than attending, but the lecture slides from the presentations look super interesting. Judging from the slides, this lecture must have been particularly awesome, though the slides from the Sutton lecture are interesting as well.

Here is the link to the schedule of speakers, along with the slides that have been posted so far[1]. If anyone in attendance has comments about how the series has been in person, I'd be quite happy to hear details.

[1] - http://www.cs.utexas.edu/mlss/schedule

Reminds me of a lyric from an old rap song. "Drive slow, homie". Ready and eager for downvotes on this unproductive comment :p

Oof. Had a hilariously similar experience doing data entry at my first internship, which was at a financial services firm.

I knew nothing about programming, but I had taken an intro stats class that had a few assignments in R. A super simple R script did half of my day's work in seconds (plus some debugging because of course I did next to no error handling).

In retrospect, I've realized that if I knew a thing of two about regular expressions at the time, I would have been able to automate my whole summer internship.

So it goes.