3 ms·
Sure, but the small & efficient networks so far have been pretty poor. Edge AI was hugely hyped a few years ago and it petered out. All of the big tech companie
by datpiff 4y ago
Sure, but the small & efficient networks so far have been pretty poor. Edge AI was hugely hyped a few years ago and it petered out. All of the big tech companies seems to be chasing the biggest networks with the largest training sets, that seems to be the direction for now.
There could be huge efficiency savings within the implementations but when training costs are already this high it seems naive to think the low-hanging fruit is still there.
- saiojd 4y agoI wouldn't be so confident. Flash attention is a recent, significant improvement to training times and sure looks low-hanging now. I was not familiar with Edge AI, interesting concept. I feel like improving the efficiency of very large models is much more likely. The recent successes will lead to an even larger influx of $ in the short term --- this is an existential threat to Google, after all. We will see where things go!