Do deep learning classification on the CIFAR-10 database with VGG-like structure approach 93% Test accuracy.
It's referenced the structure of Keras code sample of CIFAR-10.
- Test Accuracy : ≈78%
Next, let's referenced the tutorial from Jason Brownlee PhD, who is a professional developer and machine learning practitioner.
- Test Accuracy : ≈89% (Train / Valid : 94.60% / 89.35%)
Ref. https://arxiv.org/abs/1412.6071
Due to VGG-16 or ResNet-50 are so giant and deep neural network, I tried fractional max-pooling after read the research paper to make a deeper VGG-like neural network but smaller and shallower than VGG-16 and ResNet-50. To reduce overfitting, we use global average pooling layer instead of full connection layer.
- Test Accuracy : ≈93% (Train / Valid : 97.94% / 93.55%)
| VGG-16/VGG-19 | VGG-like + FMP | |
|---|---|---|
| total params | 10M+ | 1.5M |
| model size | 200+ MB | 12.4 MB |