e channels, retr…alf of the channels, retrain, and there — it runs exactly at twice the speed and works a bit worse. It’s magic. Conversely, if you happen to get more data/compute, you can immediately make your program work better just by adding more channels and retraining.
Magic = it’s just function approximation with less variables and higher bias.
If you get more data… Again, who’s going to get more data? The usual suspects. Meaning that, the same network will be ridiculously accurate only at Google’s, Facebook’s, and maybe Tesla’s. That same network would be delivered to the community who could do nothing without high quality data to continue training or retrain from scratch. It would have been nice/fair to add a paragraph on this point.
If Software 2.0 is what people are looking at, we seriously need to sit down and discuss about data, not algorithms.