I really enjoyed this article! The author walks us through design of an ML system and typical problems in a fun and understandable way.<p>The conclusion is pretty wrong though.<p>> Regardless of how complicated your program’s behavior is, if you write it as a neural network, the program remains interpretable. To know what your neural network actually does, just read the dataset.<p>This is not true at all? That’s like saying “all binaries are open source, just read the machine code”. Most datasets contain a reasonable amount of pollution (even the author ran into this). And if you let the model train on its own output it’s pretty easy for it to cheat: <a href="https://techcrunch.com/2018/12/31/this-clever-ai-hid-data-from-its-creators-to-cheat-at-its-appointed-task/" rel="nofollow">https://techcrunch.com/2018/12/31/this-clever-ai-hid-data-fr...</a><p>Moreover current ML techniques are pretty sample inefficient. So your dataset is likely to be much, much larger than the equivalent program. Right now we haven’t developed much tooling to help you map a sample bad input or behavior back to training data that might be relevant. So “just reading the data” seems like lots more work than debugging a modern program.<p>I do think we’ll get better at debugging this stuff in time, but I don’t think it’s currently true that ML systems are simpler or more interpretable than a corresponding imperative program.
The demo feels like a dream. You try to move towards things, but they transform and you can not reach them. You can also not go back, the place you came from totally changed. Suddenly, you're somewhere else completely.<p>Awesome idea and solid article!
From the author's twitter, here are links to a download of the network and to a site that lets you view the network architecture as a flow chart: <a href="https://twitter.com/madebyollin/status/1566886407117107200?s=20&t=xDwYGHvM0sqBIlwE9CTZxA" rel="nofollow">https://twitter.com/madebyollin/status/1566886407117107200?s...</a>
This was an interesting application of neural nets and a good write up! I’m curious how a game might be made from scratch using a neural net, i.e. what would it train on? I think this would be a cool technique to have at our disposal because games implemented as neural nets might make data mining harder, so all the cool secrets in the game wouldn’t be immediately exposed upon release :)
Don't forget to try the demo right in your browser: <a href="https://madebyoll.in/posts/game_emulation_via_dnn/demo/" rel="nofollow">https://madebyoll.in/posts/game_emulation_via_dnn/demo/</a>