HNHacker News
TopNewBestAskShowJobs

smhx

808 karma · joined June 22, 2014

submissionscomments
smhx··on I Trained a Deep Q Network Built in TensorFlow to Play Atari Pong
This is simply a transpile / reproduction of the original Torch version from Deepmind, but in TensorFlow. It doesn't really do anything new or different compared to the paper by [Mnih et. al.](http://www.nature.com/nature/journal/v518/n7540/full/nature1...) .
smhx··on Ask HN: Papers you read in 2015?
For people actively working full-time in the field, some of the papers which have simple but powerful ideas, reading the paper for 2 hours (or glancing at the key diagram / formula) is enough to implement it.

For example:

Deep Residual Learning for Image Recognition

Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)

smhx··on Microsoft Research wins image recognition competition
yes. more like 5-shot hold-out, and the held-out set is pretty big as well, so more legit.
smhx··on Facebook Open Sources Its AI Hardware as It Races Google
NVIDIA M40s, which are the new cards, so double that estimate.
smhx··on Deep Convolutional Generative Adversarial Networks
The input to the G-network is a Vector of numbers, of size K. The output is an image. Different random vectors give different output images.

We are doing vector arithmetic in this input space, and the outputs seem to be semantically changing according to the vector arithmetic.

smhx··on Deep Convolutional Generative Adversarial Networks
Both use GANs to generate images, and that similarity is indeed there. Original GAN paper: http://papers.nips.cc/paper/5423-generative-adversarial-nets
smhx··on TensorFlow Benchmarks
I'm the one who runs the benchmarks. It's sad that you put such a twist to the whole thing.

I've been running convnet-benchmarks forever now, and I've been running them independently on separate personal hardware. I do this as a hobby.

I've done an apples-to-apples comparison, and my benchmark review only puts the facts forward, I dont attack them.

If you read my other social media comments, I've been pretty positive about TensorFlow, and I've even put in some groundwork to write Torch bindings for it.

Please stop spewing nonsense interpolated from like 2 super-weak data points.

smhx··on TensorFlow Benchmarks
I've done an apples to apples, TensorFlow + CuDNN R2 vs Torch + CuDNN R2
smhx··on Jeff Dean explains TensorFlow [video]
They didn't release the multi-machine version of TensorFlow. They said they're still working on it and will release it when it's ready.
smhx··on Show HN: NotCRUD – Hacker News of India
I'm not able to see the full title of the posts (13" MBP Retina). It is cut off half-way. As a front-page for a HN-style site, that's a serious usability bug that you want to fix.

http://imgur.com/DatjKpe

smhx··on Click-o-tron – Clickbait site using deep learning
Technical write-up on how this was done: http://larseidnes.com/2015/10/13/auto-generating-clickbait-w...

Source code: https://github.com/larspars/word-rnn

smhx··on Fast Convolutional Nets with Fbfft: A GPU Performance Evaluation
I run convnet-benchmarks, and also work at Facebook.

Unlike what the public perception is, NVIDIA, Facebook and other companies collaborate very very closely. NVIDIA's latest R3 release has FFT-based convolutions inspired by Facebook's work, and Facebook integrated some of the optimization suggestions for fbfft that NVIDIA gave. Further optimizations on the FFT approach (like tiled-FFTs) have been suggested and prototyped at FB and carried over by NVIDIA. The relationship is more of a deep collaborative effort.

At the end of the day, we want cuDNN to win, as it is a baseline system for all software.

However, if you want to do research, you want to have open-source code, because you often want to do funky changes to the base algorithm. NVIDIA falls short on this, and we at Facebook pledged to always keep our code open source (bitten by the fact that we had to start with closed-source blackboxes for our initial FFT work).

smhx··on Deep Learning Courses
OpenCL neuralnet backend for the Torch framework. Supports all kinds of stuff.

https://github.com/hughperkins/clnn

smhx··on A Neural Conversational Model [pdf]
Quoc Le is neither an author of the word2vec paper or in the committers list of word2vec software (there's only two committers, mikolov and sutskever). Not sure where you got that inference from.
smhx··on Lush Programming Language
I used Lush for my homework as part of the Machine Learning course at NYU (taught by LeCun). It was pretty mature, mostly bug-free for general use. However, at that time, I never used a lisp-y language before that, and found it quite difficult to do my homeworks. It had an interesting mode where if you annotate your lush source-code with type-hints, all your Lush transpiles in to C, which you can compile with your favorite C compiler. It was eventually abandoned in usage because Torch (Lua based scientific-computing platform) was adopted by the lab.

AFAIK, Lush is not used inside Facebook labs.

smhx··on Offer HN: I'll develop your mobile app at a bargain basement price
Can you name a few such outsourcing shops that seem too good to be true (and likely are).
smhx··on Facebook's Memory Networks Demo
Memory Networks: http://arxiv.org/abs/1410.3916
smhx··on Show HN: Let's Chat, a self-hosted chat app for small teams
For github projects, I've been using http://gitter.im I've found the integration to be awesome and well thought-out.
smhx··on Facebook open-sources deep-learning modules
there's clear and well stack-traces at the lua level. there's great debugger (in fblualib, called fb.debugger). there's an inbuilt profiler into LuaJIT. too bad you've had a bad experience.
smhx··on Deep Image: Scaling Up Image Recognition
googlenet already has 5.5%, they published it at a bay area meetup, but did not officially publish the numbers yet!
smhx··on Nvidia Introduces CuDNN, a CUDA-based Library for Deep Neural Networks
With the power-draw of those things, I think it would be a little tricky in the short term for tablets (and even harder for phones)
smhx··on Nvidia Introduces CuDNN, a CUDA-based Library for Deep Neural Networks
In these deep relu networks which are not renormalized in between, (like overfeat which has no normalization), some of the activations become pretty big in size! (in the order of 1e3).

You also cant clip them and get away with it, you have to either renormalize the layers to do half-precision (and live with the extra cost) or stick to full-precision. I was doing fun stuff early this year that did fixed-precision nets (8-bit/16-bit). Things get very interesting :)

smhx··on Nvidia Introduces CuDNN, a CUDA-based Library for Deep Neural Networks
meta-optimizers need to factor in a few more things like memory usage. Right now the FFT modules are insanely memory-hungry, and the versions written with reasonable memory usage (Michael Mathieu wrote one), are not as fast as batched CuFFT + CuBLAS
smhx··on Beaker Notebook v1.0 released – open-source polyglot data science web app
Can someone familiar compare and contrast it with iPython Notebook (which also has support for any language/package that writes an ipython kernel plugin)
smhx··on [dead]
I like square cash a tad more! It supports any platform (by design) and it works great for me.
smhx··on Show HN: Purzue – Résumés redefined
I dont mean to hijack the post, but I recently found LinkedIn's resume generator

http://resume.linkedinlabs.com/

and I love it. Would Purzue have a plugin to import your LinkedIn profile or Github profile for example?

smhx··on Torch7 – Scientific computing for LuaJIT
my biggest annoyance with LuaJIT is the lack of 64-bit addressing, and there's no easy fix for this. Mike Pall said he'd fix it in LuaJIT 3.0 (with an overhaul of the GC), whenever that'll be.

http://hacksoflife.blogspot.com/2012/12/integrating-luajit-w...

Edit: to clarify, this does not mean torch cant use more than 2GB, but native lua allocations cant be more than 2GB, so ffi-based or C based allocations (which is how torch tensors are allocated) have no such limits.

smhx··on Torch7 – Scientific computing for LuaJIT
Someone I met recently said this (in rough words): "When I'm writing rough code in lua, it's completely acceptable to do a couple of for loops here and there without a disaster in speed. With python, this was a complete meltdown"

I'm not saying it all boils down to just this quote, but I thought it was interesting and wanted to replay it here.

smhx··on Torch7 – Scientific computing for LuaJIT
go for it.
smhx··on Torch7 – Scientific computing for LuaJIT
One thing that isn't really supported yet :-( We have support for doing complex FFTs etc. in the signal package, but otherwise, it is non-existent.
← PreviousPage 3 of 4Next →