Backpropagation-free AI training accelerates 4000x in 10 days to 1.6 seconds
Original article in Korean역전파 없는 AI 학습 실험, 열흘 만에 4000배 빨라져…1.76시간→1.6초
- Outlet
- AI Times
- Published
- Importance
- 3/5
Summary
An open-source project called Noise demonstrated backpropagation-free language model training reaching target performance in 1.6 seconds, down from 1.76 hours in 10 days. The open-source community improved the training algorithm and GPU computation methods, achieving a 4000x speed improvement.
#optimization#backprop-free#learning
Read the originalaitimes.com · in Korean