transcranial

Looks really amazing! Reproducibility of models and experiments is huge. It should be almost a requirement if one is going to publish results claiming SotA, etc. Seems like you could become the GitHub for deep learning in addition to the Heroku for deep learning.

I've been using Stylish (https://userstyles.org/styles/browse/github) for this -- works pretty well.

Unfortunately the paper is in Nature, paywalled, instead of Arxiv, and data/code/model/weights inaccessible. While publishing in Science/Nature/NEJM/JAMA is definitely the right approach for deep learning to gain validity in the medical community, faster progress could be gained by having a more open platform, with constant and real-time validation with more data, more medical centers and clinics. The reason progress in DL has been so breathtaking is in no small part due to the culture of openness and sharing.

I think medicine will over time morph from a field generally perceived as an intellectual one, to a largely humanistic one, like nursing. Most people, especially the HN crowd, vastly underestimate the importance of the human touch. A doctor who possesses this can make a huge difference, but sadly the are outnumbered by those who do not.

I actually played the amazing game Monument Valley after seeing it in House of Cards. Although from what I can recall this wasn't paid placement so much as the writers just loved the game.

They _do_, but they _shouldn't_. It's an inefficient allocation of skills and resources.

Have you seen Keras.js? It uses WebGL to run certain computations on the GPU. https://github.com/transcranial/keras-js

Yeah, setting aside whether his claim is actually true or not, why would this even be newsworthy or matter, if it were not for this context?

Thanks. Nothing fancy with the network architecture diagrams. The layers are just div elements with the layer name as the id. There's a definition of inbound/outbound connections by layer names, extracted from the keras json config, which is used to draw SVG paths.

You mean with tensorflow or theano as the backend? They have all kinds of optimizations that isn't possible to replicate here yet. There is certainly room for optimization! Also, 1000-layer resnets should theoretically be possible, but probably isn't that practical. Lots of exciting work happening in searching for more efficient architectures.

Thanks! For WebGL, credit goes to https://github.com/waylonflinn/weblas. I only really use GEMM, but it works quite well. In keras.js, convolution is implemented with the oft-used im2col transformation to turn it into a matrix multiply followed by reshape. Convolution kernels directly GLSL could potentially provide speed gains I'm sure, but I can't even imagine writing it for tensors of arbitrary shape.

If anybody likes transparent backgrounds, I wrote up a quick guide on how to enable it for Atom: https://github.com/transcranial/atom-transparency

And https://arxiv.org/abs/1603.07285 for the corresponding paper. Really clear and easy-to-understand explanation of some of the math.

Interesting that the base data consists of URLs. I guess it makes sense given copyright issues. Anybody know what the ballpark expected half-life of such URLs?

Agreed! In many ways it is massively underrated precisely because it isn't backed by a Google or Facebook. Definitely glad that people are seeing its awesomeness based on merits. I love React and its whole ecosystem, but Vue definitely stands on equal footing. There's good reasons to love both.

Pretty sure he means anti-infectious disease

Nice work! I've also tried my hands at use WebGL to do deep learning in the browser (e.g. https://github.com/scienceai/neocortex/blob/master/src/lib/w...). The conclusion I came to was that there are just way too many limitations for it to really pay off. The need to encode everything in textures etc limits the data shape and dimensionality, not to mention the complexity cost. If you can get more complex networks working I'll be really impressed!

MXnet.js [1] is an emscripten port of the base C++ framework. It runs entirely in the browser and works fairly well. The actual code produced by emscripten isn't that large, but the model weights can become an issue. I've tried to get emscripten working on tensorflow, even just for forward prediction, but have been pretty much gotten nowhere. Of course this doesn't let you harness GPU power.

Lots of cool potential applications of doing deep learning over the web are just waiting to be discovered and built.

[1] https://github.com/dmlc/mxnet.js

Nice list here, but by no means exhaustive: https://github.com/h5bp/Front-end-Developer-Interview-Questi...

In my experience, the only time this happens is with `npm uninstall`. If you only `npm install` stuff forever, node_modules seems pretty stable. Disk usage, however, is another matter.

The problem is that the latent space of a model trained on all comments is still massive. For any sort of coherence, you need to condition the probabilities on smaller spaces, or enforce structural bounds.

Here's an example of a paper where they simultaneously optimize high-level semantics with token-level sequence probabilities: http://arxiv.org/pdf/1606.00776v2.pdf.

Yeah, there's no effort to direct or encode intent. It's just producing what it probabilistically wants to produce. If it happens to run...cool?

Here's a interesting example from Google DeepMind learning to generate code for Magic The Gathering and Hearthstone: http://arxiv.org/pdf/1603.06744v2.pdf

Apparently Parkinson's disease has been on the rise over the past several disease, some say due to the decrease in smoking. Which is fascinating, as smoking has been shown to be actually protective against Parkinson's.

Currently, South Korea and Japan have the highest incidences of stomach cancer. It's typically attributed to cured foods and processed meats. Prior to refrigeration, wide consumption of these foods were out of necessity.

The correlation is one of the strongest there is, and indeed, causal -- there's a almost perfect lag time of ~20 years [1].

[1] https://en.wikipedia.org/wiki/Lung_cancer#/media/File:Cancer...

Interestingly, this paper cites that previous submission:

"2. Dahl, R.: Automatic colorization. In: http://tinyclouds.org/colorize/. (2016)"

They even included it in their comparisons (Table 1; Figure 5).

Congrats! Atom may not be perfect (no editor is), but it's constantly and consistently improving, which is a lot more than one can say about most other editors.

For those who use Jupyter: there's a package called Hydrogen that let's you run your code directly within Atom using Jupyter kernels (not just python!). To me, that's the perfect example of the power of a hackable editor.

They say that they make use of GPUs. Anybody know if this means GPU VMs are coming to google compute engine? That would be amazing. AWS GPU instances are popular but somewhat limited as they use older hardware.

Might be interested to check out https://science.ai

Nice work! Another demonstration of the amazing power of deep neural nets for transfer learning. For those interested, http://arxiv.org/abs/1511.06681 demonstrates experiments with video coloring using a 3D convolution-deconvolution network with some architectural elements similar to that used by the author.

Really cool insight / results. Like relu and dropout, love it when such simple techniques make such great improvements.