A study examining the dependence of neural networks on initial values in a simulation called Game of Life. Game of Life can be represented by a two-layer CNN because it uses relatively simple rules for updating, but in reality, the rules can only be learned correctly in a much larger network, and the larger the It learns robustly to initial perturbations. The experimental results support the lottery ticket hypothesis.
TL;DR
A study examining the dependence of neural networks on initial values in a simulation called Game of Life. Game of Life can be represented by a two-layer CNN because it uses relatively simple rules for updating, but in reality, the rules can only be learned correctly in a much larger network, and the larger the It learns robustly to initial perturbations. The experimental results support the lottery ticket hypothesis.
Why it matters:
Paper URL
https://arxiv.org/abs/2009.01398
Submission Dates(yyyy/mm/dd)
2020/09/02
Authors and institutions
Jacob M. Springer, Garrett T. Kenyon
Methods
Results
Comments