iverjo

Sadly, pip 20.3 seems to have broken docker builds in one of my projects. The symptom is that the pip install seems to hang indefinitely (>40000 seconds). I switched back to 20.2 for now.

Thanks! No, it doesn't, but I'm happy to accept a pull request for such an option

Shameless plug: Here's my take on building images with images. It uses a genetic algorithm. https://github.com/iver56/emoji-art-generator

Further, you could shave off a couple of bytes by inserting

  <body onload=F()>

before the canvas element, instead of having

  <script>F()</script>

after the canvas element

I think you could shave off 8 characters by doing

  0|Math.random()*3

instead of

  parseInt(Math.random()*3)

I also had that idea, and I made a crude proof of concept: https://github.com/iver56/cnn-clock

This idea originated in the jsgolf community and dwitter (a site where you post 140 character js snippets that generate interesting visuals). See https://www.dwitter.net/u/aemkei

I spend almost all time programming and thinking. Occasionally, I discuss some issues with colleagues. Other than that, I attend approximately one hour of meetings per week in total. I like this hands-off kind of management.

A friend of mine once wrote a 3D effect in a 2D canvas in only 130 bytes of JavaScript code: https://www.dwitter.net/d/914

Sony has been working on style (genre) transfer in music lately: https://youtu.be/I9M8l2guPSo?t=21m30s

Probably to avoid confusion, as git (and github) already had a concept of tags

https://help.github.com/articles/working-with-tags/

Do you mean messages sent over infrasound (less than ~20 Hz)? That would typically not work, because most speakers cannot produce infrasound and most microphones don't pick up frequencies that low

This is very cool! Kudos to the authors. I had the idea about this about a year ago (when style transfer for images was gaining traction), and wanted to do it in my master's thesis. I ended up doing something related, though: Evolving neural networks that transform sounds with the help of audio effects. For example, they can process white noise so it sounds like a drum loop. Sounds are available here:

http://crossadaptive.hf.ntnu.no/index.php/2016/06/27/evolvin...

How does this relate to Progressive Neural Networks [0]? That technique is also about accumulating knowledge (while not forgetting existing knowledge)

[0] https://arxiv.org/abs/1606.04671

Jokingly, this was the first apple AI research I saw online, about a month ago: https://arxiv.org/abs/1610.08120v1

To the author: Have you tried to use a logarithmic frequency scale in the spectrogram? [1] That representation is closer to the way humans perceive sound, and gives you finer resolution in the lower frequencies. [2] If you want to make your representation even closer to the human's perception, take a look at Google's CARFAC research. [3] Basically, they model the ear. I've prepared a Python utility for converting sound to Neural Activity Pattern (resembles a spectrogram when you plot it) here: https://github.com/iver56/carfac/tree/master/util

[1] https://sourceforge.net/p/sox/feature-requests/176/

[2] https://en.wikipedia.org/wiki/Mel_scale

[3] http://research.google.com/pubs/pub37215.html

Nice approach, and well explained! By the way, Niland is a startup that also does music labeling with the help of deep learning.

Demo available here: http://demo.niland.io/

For example, it can output Drum Machine: 87%, House: 88%, Female Voice: 55%, Groovy: 93%

It's also inspiring to see how much you can do with only 140 characters of code (so it fits in a tweet). Here are some examples:

https://www.dwitter.net/d/104 (tunnel)

https://www.dwitter.net/d/302 (old school effect)

https://www.dwitter.net/d/406 (rotating cylinder)

I recommend you to try the LAB color space and to use delta E 94 [1] for computing distances between colors in the color space

[1] https://en.wikipedia.org/wiki/Color_difference#CIE94

I'd say scikit-learn is a better way to get started with machine learning. Check this out, for example: https://www.youtube.com/watch?v=cKxRvEZd3Mw

I guess one could train a Convolutional Neural Network to convert the ASCII art to an actual image. Something like this, but for ASCII: http://sketchy.eye.gatech.edu/

Wow man, sigvehtml is actually used in production!

Source: https://github.com/sigvef/arktis/blob/master/index.sigvehtml

Rendered website: http://arkt.is/

We don't have a blog yet. We're using Azure for the hardware and mainly scikit-learn for the training (we train only on metadata at the moment). Will probably start using Tensorflow soon.

This is nice :) Kudos to the Youtube guys for releasing this. I'm a data scientist in a startup where one of the things I do is create multi-label models for classifying YouTube videos. My current model has 90 % precision and 69 % recall, while Youtube-8M has 78 % precision and 14 % recall, with respect to the human raters. I guess one of the reasons is that my model only has around 100 categories, while Youtube-8M has 4800. It's like comparing apples with pears, but still interesting.

Fractals can be drawn by surprisingly small amounts of code. For example, this mandelbrot fractal is implemented in 122 characters of JavaScript: https://www.dwitter.net/d/123

I wouldn't be surprised if the burning ship fractal could be coded on dwitter (i.e. with 140 characters or less) as well

This brings Lime [1] to mind. "Explaining the predictions of any machine learning classifier"

[1] https://github.com/marcotcr/lime

In one of my web app projects, I went the other way. I converted SVG icons to appropriately sized PNGs. That was after I found out that SVG rendering was slow on mobile, and was the culprit of janky scrolling.

Would love to see some more details on how you represent videos as feature vectors. Do you only use metadata provided by the uploaders (e.g. title and tags), or do you also analyze the raw video/audio somehow to augment the metadata?