Hacker News new | past | comments | ask | show | jobs | submit login
My phone case, the hidden state of an RNN (douglasorr.github.io)
71 points by douglasorr on Sept 7, 2021 | hide | past | favorite | 36 comments



Beautiful (and I love the hidden nerd-cred). In the late '90s I bought a PCB briefcase in Germany and thought I'd reached the pinnacle of nerdom but this is clearly better.

- https://www.worthpoint.com/worthopedia/slim-briefcase-291857...


I have a whole toolbox made of PCBs. That slim briefcase looks much more convenient.


Don't know what security on airports would think about that one today, heh.


I have carried mine without issue.


My desktop background is the output of a neural network with random parameters that maps (x, y) to (r, g, b): https://i.imgur.com/KySV77c.jpg.

Here's a small album of other wallpapers I've made like this: https://imgur.com/a/BrNhwkF


"This post may contain adult or erotic images...". Well I didn't knew neural network were in the nsfw categories now.


Related.../u/gabrielgoh in 2016 turned Yahoo's nsfw neural network into a weird abstract image generator. The gitlab repo is private now but you can read some of the comments https://www.reddit.com/r/programming/comments/58k3sf/image_s...

https://news.ycombinator.com/item?id=12756462

Here's a blog post on it, and a warning, it's perceptibly NSFW

https://www.gwern.net/docs/ai/2016-goh-opennsfw.html


> The gitlab repo is private now

The internet always remembers:

https://archive.is/20181205151107/https://open_nsfw.gitlab.i...

https://archive.is/20170506185728/https://open_nsfw.gitlab.i...

(But seriously, that stuff is NSFL. I want to bleach my eyes now. A lot of those are like erotic lovecraftian fantasy landscapes.)


Looks great but I'm surprised they aren't way noisier. Is that just a result of having a limited number of neurons compared to the number of pixels, or is there some specific trick to it?


I think at least some of the structure is down to the update equation, roughly 'new = sign(gate * old + ...)'. Since 'gate' is always positive, this means that even when the network is randomly initialized, the state is more likely to stay the same between two time steps than to change.

I tried "deep transitions" which repeats this equation a couple of times, and the results were indeed noisier.


It's mostly that these are very narrow and shallow networks -- if I remember right, 5 layers and 30 units wide?


Not OP, but I imagine you can set maximum limits on the rate change of the RGB vector so it doesn't create a static but a more fluidic appearing image.


It just occurred to me that a phone case is probably a really good place to put some adversarial pixels.


Do general purpose adversarial patches even exist? I thought they were pretty specific to the particular network, subject and viewing angle?


they do need to be tuned at least somewhat.


What kind of adversarial pixels do you mean? Those that confuse AI object recognition software?


Yeah, I had the same thought — I'd much rather have an adversarial patch on my phone case than this.


It might work to reorder using some heuristic based method similar to the AMD methods for matrix factorisation reordering which is used to minimize infill for sparse matrices.

Regardless, cool project!


I've wanted to try out sorting neural net dimensions for visualizations via a travelling salesman problem solver over the correlation matrix among the dimensions. Has anyone tried something like this?


Nearly indistinguishable from noise. Subtle art, and hidden meaning. Inspiring stuff, though I can't help but wonder if noise + a stupid algorithm couldn't be even more interesting.


Very cool! I love generative art, and especially when it's backed by some cool concepts.

Yesterday I made this https://i.imgur.com/MqJGt4P.png Abstract image, but it's a circle packing algorithm (borrowed from d3) running on hierarchical data where I work. Thought it looked cool.


cool! Yours is rather mesmerizing to stare into :)

In a somewhat similar vein, I actually just finished a weekend generative art project[1] inspired by the recent HN post on minsky circles[2].

1 - https://wcarss.ca/circles/

2 - https://news.ycombinator.com/item?id=28368626


Yeah, the full size is almost 200 megapixels. Would be cool to have a wall full of these not-random-but-different holes.

Enjoyed yours as well. Perfect amount of random. From the code it looks like you've tweaked a lot. Good job!


Looks so much like cellular automata! I wonder what initial conditions would output that :)


For a second, I was hoping the image on the phone case was an adversarial one, intended to jam up facial-recognition other surveillance-related AI. No, nothing so extravagent, just a visualization of internal state.

'Tis still nifty.


The idea is pretty cool, and well executed.

I do think the coloring is what makes it jarring and ugly though. They could’ve gone with subtler, mixed colors - instead of a bright uniform blue.


I’m pleased he has made an artsy phone case he likes really, but this is another round of Raspberry Pi feeblemindedness. Why would someone with a powerful skill in recursive neural networks make a phone case? - when there are so many areas to focus on, challenges we’re facing together, where even failure would deliver real value to the community. What’s up with these hobbyists and their cute blinking LED projects? You don’t have to change the world but choose some objective that at least has a chance of delivering real value and utility. What am I missing? Tell me I’m wrong and how


I mean he enjoyed making it! Not only is he increasing his skills making his next project better he is also sharing it with others potentially getting more people interested in that field both of which are helping the "community"!

Really whats the point in living if all you do is work for the "grater good" without ever having fun? I got into programing purely because I wanted to create something for a video game that was missing. In zero way was this productive for society! I'm 99% sure that Doug got into programing for a similar reason: because it was fun for him!


Are people not allowed hobbies? Are we only allowed to do both what we are good at and what will have the most impact?


Well he’s free to do whatever he wants. But creating someone with that education is a type of investment society makes. There’s a high cost to the community and yes of course the group expects something in return. It’s kind of like the lying flat movement in China. They can do what they please, but wouldn’t it be nice if they enjoyed helping the people around them? Is that too much to ask?


Working 40 hours a week is payback enough, what people do with their other time is their decision. I hate the need to always hustle or be productive.

Besides, he made something many here enjoy. Thus his work did have value.

If anything, allowing people to experiment is probably a net win for society in the long run.


Yes it is too much to ask. We cannot turn people into robots that do our bidding.


But isn’t that exactly what we are? Complex biological machines designed by the group through natural selection manufactured in volume in a distributed factory with the express and only purpose to serve our creator - since we were made in its image. (I’m atheist but that fit so well)


No we are not machines. You can tell a computer to try and solve a problem for years on end without a problem. Try and get a person to spend their entire lives trying to "improve" the world without any diversions or fun in the slightest and they'll burn out within a week.


Why are you wasting the investment that your community made in you, by spending time trolling on hackernews? You’re not helping further anything for society with these comments.


Crossing my eyes I can almost see the dinosaur ....




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: