Summary
Research Objective
There are recent found that, from a large neural trained network, we can prune and obtain a small sub network (even 90% of the parameters is being pruned), without compromising the performance. It natural to think that, if we could have a way, train a small network from the scratch, also obtains similar performance as the large network, saving the energy for training. According to current experience, a pruned sparse network is hard to train from start.
Problem Statement
The author proposed the Lottery Ticker Hypothesis, stating:
A randomly-initialized, dense neural network contains a subnetwork that is initialized such that—when trained in isolation—it can match the test accuracy of the original network after training for at most the same number of iterations.
Methods
The training is procedures are
- Init a large dense network, init with $\theta_0$
- Train the network
- Find a mask, prune with Least Weight Magnitude
- Rest the weight to $\theta_0$, and apply the mask
Repeat above steps, prune iteratively.
Contribution
- The interesting finding of the lottery ticket hypothesis, that existence of sub network reach better performance and higher accuracy than the original network within fewer iterations. The case under the constraint that the parameter should be the same as the initial param of untrained dense network.
Evaluation
Conclusion
Notes
References
Background:
- Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXivpreprint arXiv:1503.02531, 2015.
- Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in neural information processing systems, pp. 1135–1143,2015.
- Tien-Ju Yang, Yu-Hsin Chen, and Vivienne Sze. Designing energy-efficient convolutional neural networks using energy-aware pruning. arXiv preprint, 2017.
- Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. Thinet: A filter level pruning method for deep neural network compression. arXiv preprint arXiv:1707.06342, 2017.