2 ms·
New training techniques are released on a monthly basis, but the basic structure of convolutional neural nets has actually been around since the '90s and hasn't
by modeless 11y ago
New training techniques are released on a monthly basis, but the basic structure of convolutional neural nets has actually been around since the '90s and hasn't materially changed. I think that's a stable enough target to build an ASIC for. Even if you're unable to implement the year's hottest training technique for some reason, if you're 10x bigger/faster you'll likely still set world records using last year's techniques.