Icml2022 | utility theory of sequential decision making

Thesis link ：https://arxiv.org/pdf/2206.12562.pdf

Large based on Transformer The model has shown superior performance in various naturallanguageprocessing and computer vision tasks . However , These models contain a large number of parameters , This limits their deployment in real applications . To reduce the size of the model , Researchers prune these models according to the importance score of the weight . However , These scores are usually estimated in small batches during training , Due to small batch sampling and complex training dynamics , This brings a lot of variability / uncertainty . Because of this uncertainty , Common pruning methods prune some key weights , Make training unstable , It is not conducive to generalization . To solve this problem , We proposed PLATON Algorithm , The algorithm uses the upper confidence limit of importance estimation (upper confidence bound, UCB) To capture the uncertainty of the importance score . Especially for the weight with low importance score but high uncertainty ,PLATON Tend to keep them and explore their capacity . We are in natural language understanding 、 Question answering and image classification are based on transformer A large number of experiments have been carried out on the model , To verify PLATON The effectiveness of the . It turns out that , At different sparsity levels ,PLATON The algorithm has been significantly improved .