Researchers at MIT created CompreSSM, a method that shrinks AI models while they train instead of waiting until after.
← News
New technique makes AI models leaner and faster while they’re still learning
Up To 4x Faster Training speedup
85.7% (Vs 81.8% Small-From-Scratch) CIFAR-10 accuracy
128→12 Dimensions Compression
Key Parts Found After Just 10% Of Training Efficiency insight
3.9 Points accuracy edge
4x training speedup
10x compression ratio
10% training completion