nicklo

I wonder what this means for computer vision startups like http://clarifai.com?

Its hard to compete with Google on a task like image classification when Google has immense computational resources, tons of data, and hoards of top researchers.

Terrific post! Its great to see things implemented in short, 1-file programs. Tends to demystify things a lot.

What did you use to make the diagrams + gifs?

Not sure I agree with "hastily launched". First public commit to the github repo was January 14th 2014.

Link: https://github.com/Samsung/veles/commit/7b19147069119399c89f...

The presumption that this or Google's Tensor Flow can be cobbled together in a few hours is completely absurd. These things take months if not years of development. And releasing them into the wild is not a process that takes 30 minutes. Open Sourcing a company library is a work intensive process that takes weeks if not months. Beyond the code itself there are a swatch of promotional materials to prepare.

That said, I do agree that the submission itself to HN was in response to Google's Tensor Flow announcement. The submitter created their acct an hour ago and the only other activity is a comment on the google announcement talking about veles.

Also the submission title was unnecessarily snarky.

Yann gave a talk at my university last week. It's really astounding how far deep learning has come from performing classification and regression to now doing things like reasoning and generation. Does anyone else find this wild? These are tasks that historically required a tremendous amount of domain-specific engineering to perform well. And recurrent memory networks just power right on through them.

I understand that learned, dense vector representations are powerful blocks to operate on, but hot damn the trend of general systems that require less domain-specific engineering is an exciting one that blows my mind.

Glad to see a critical eye on these sorts of things. I had a question about the choice of loss function. So regression is definitely the wrong choice since our outputs are binary (0 or 1) and not real-valued (0-1).

We could use a softmax classifier as this is meant for multi-class binary classifications. Each class in this case would represent a neighboring pixel, a classification of that class would represent activating that pixel. However, softmax assumes one-label and the probabilities add up to 1. Our problem is a multi-label multi-class binary classification.

We could train 9 of such softmax node groupings for each neighboring pixel but that immediately seems to be a bad idea.

Another awful solution - make each possible pixel configuration (2^9 of these) a class and do a standard soft-max.

I can start to see why the author chose to use regression loss as a sort of hack to get this to work, but I'm trying to think out the best, proper solution.

Any thoughts?

This was the premise of a South Park episode last year.

It's amazing that Trey & Matt are so dead on when it comes to current events and culture critique that they sometimes even parody the future.

I think I won the game

http://imgur.com/qSerUAV

The note text box breaks iOS autocomplete support. It's doing something funky where the keyboard api can't recognize previous words. May want to consider just using a native text box instead of something custom and funky with JavaScript. Will probably resolve the typing performance issues others have mentioned.

Hey thanks! This was super useful.

Google has published a lot of work recently about using deep learning for language translation with really impressive results. Anyone know if Google Translate is now neural net based? Or would a many-layered net not scale to the amount of API calls that Google translate receives?

Yeah I wonder what the false positive rate is. Didn't see any mention of it though.