Glad you dig it. Although explorations like these are still kind of just toys, we think they begin to inform how ML will impact design practice.
We tried training on bitmap pixel data, but our snapshots were pretty big making for troublesome training arrays, even when down-sampled… training took forever. We were also interested in getting some experience training vectors for some future explorations we have in mind that have to do with optimizing positioning of shapes. Haven’t been able to wrap our heads are ConvnetJS yet — digging into Synaptic.js now.
Thanks for your note.