Jeffrey Stone
I design high speed datapaths for a living: chips that encrypt, protect, compress, detect and calculate. Why would I then spend evenings and weekends designing labs and applications to give away for free? Quite simply, I am giving myself the freedom to follow wherever my curiosity leads. I am publishing the results in hopes that others will share that same joy of discovery.
The game
iOS & Android BetaThe October 1977 issue of Byte magazine arrived when I was a teenager building my own computer. Its cover theme was "Implementing Space War", and there on page 86 was an article by Dave Kruglinski, "How to Implement Space War (or Using Your Oscilloscope as a Telescope)". Everything about the article fascinated me, starting with the fact that a '70s era microcomputer, 1/1000th of what we carry in our pockets today, could make the calculations for a physically realistic orbital mechanics simulation.
I have written an almost embarrassing number of Spacewar games in the years since, tracking Moore's law with more and more sophisticated versions. I developed a ladder of 20 programmed levels. The game became more fun and more challenging, and looked great with some aesthetic flourishes as a mobile app. Then, I got curious about machine learning. The spine-tingling excitement of building a system that does not need detailed programming, but can learn to play a video game by trial and error, is frankly addictive, and it has led to so much more. For a start, Spacewar: Orbital Duel is based on a hybrid AI opponent, a neural pilot with a few scripted guard rails. Your opponent learned by playing. I never told it how.
The labs
The neural network that you are playing against in Spacewar: Orbital Duel uses reinforcement learning in a deep neural network. This technology broke out in 2006 to 2012, but is based on concepts that date back to the 1940s and 1950s.
The Spotter lab can play Spacewar from nothing but what it sees on the screen, with a convolutional neural network that does image segmentation. It is based on the 2015 "U-Net" paper (Ronneberger et al.). CNNs took over image processing around 2012, built on ideas worked out in the '80s and '90s. The Spotter is rather contrived. I am working on a much more complex CNN lab that can identify rivers, and distinguish where the rapids are, from an aerial photograph.
Rabbitbrush: Plant ID brings the ML story very near to the present day. It relies on a transformer (the T in GPT) with an attention mechanism to build a network that is capable of distinguishing 4,000 plant species by their leaves, buds, flowers or bark. The Transformer paper, "Attention Is All You Need", was published by Google researchers in 2017. ChatGPT brought the approach to the mainstream in 2022.
Four nets, four orders of magnitude
Parameters per network, log scale. Largest ÷ smallest ≈ 18,000×.