Statements (30)
Predicate | Object |
---|---|
gptkbp:instanceOf |
scientific theory
|
gptkbp:author |
Jonathan Frankle, Michael Carbin
|
gptkbp:citation |
gptkb:ICLR_2019
gptkb:The_Lottery_Ticket_Hypothesis:_Finding_Sparse,_Trainable_Neural_Networks 2019 |
gptkbp:debatedBy |
applicability to transfer learning
effectiveness in different architectures generalizability to large-scale models |
gptkbp:field |
gptkb:machine_learning
|
gptkbp:hasConcept |
Dense, randomly-initialized neural networks contain subnetworks that can be trained in isolation to reach comparable accuracy to the original network.
winning ticket |
https://www.w3.org/2000/01/rdf-schema#label |
Lottery Ticket Hypothesis
|
gptkbp:influenced |
network initialization studies
pruning algorithms research on efficient neural networks |
gptkbp:method |
Iterative pruning and retraining
|
gptkbp:proposedBy |
gptkb:Jonathan_Frankle
gptkb:Michael_Carbin |
gptkbp:publishedIn |
gptkb:International_Conference_on_Learning_Representations_(ICLR)
|
gptkbp:relatedTo |
deep learning
model compression neural network pruning |
gptkbp:testedBy |
gptkb:ImageNet_dataset
gptkb:MNIST_dataset gptkb:CIFAR-10_dataset image classification tasks |
gptkbp:winningTicketDefinition |
A sparse subnetwork that can be trained from its original initialization to match the accuracy of the full network.
|
gptkbp:yearProposed |
2018
|
gptkbp:bfsParent |
gptkb:Jonathan_Frankle
|
gptkbp:bfsLayer |
7
|