I've trained two neural networks to communicate classes (based on images) across a noisy channel. One constraint is they must compress the information into a lower-resolution image.

Additionally, they're tasked with fooling another network into choosing the wrong class based on their images. It's like having a couple whisper in code while someone nearby eavesdrops.

The idea is to build on what people have done with creating new alphabets by encoding and decoding data with two NNs. What's new here is that the communication is "about" some stimuli and flows through the same medium as the stimuli.

Also, the NNs are tasked with playing a sort of game of telephone, passing messages about the same stimuli back and forth, adding noise as it goes.


The whole thing is loosely influenced by Wittgenstein's philosophy surrounding "language games."

Sign in to participate in the conversation
Refactor Camp

Mastodon instance for attendees of Refactor Camp, and members of various online/offline groups that have grown out of it. Related local groups with varying levels of activity exist in the Bay Area, New York, Chicago, and Austin. Kinda/sorta sponsored by the Ribbonfarm Blogamatic Universe. If you already know a few people in this neck of the woods, try and pick a handle they'll recognize when you sign up. Please note that the registration confirmation email may end up in your spam folder, so check there. It should come from administrator Zach Faddis.