This project uses Evolution Strategies on mnist task.
The structure of the network is followed by here.
| accuracy | Layer number | iteration times | npop | optimize method | GPU/CPU |
|---|---|---|---|---|---|
| 92.05% | 1 | 1000 | none | BP | CPU |
| 99.11% | 3 | 20000 | none | BP | CPU |
| 90.94% | 1 | 20000 | 50 | ES | CPU |
| 91.31% | 1 | 20000 | 50 | ES | GPU |
| 83.32% | 3 | 10000 | 10 | ES | CPU |
| 85.8% | 3 | 10000 | 20 | ES | CPU |
| 85.53% | 3 | 20000 | 10 | ES | CPU |
| 82.5% | 3 | 10000 | 50 | ES | CPU |
| 84.19% | 3 | 10000 | 10 | ES | GPU |
The writer says ES method can get better results than BP (backpropagation) method.
But I results showed that it is NO, I don't know where is the problem.