CNTK image classification result improvement method

Question

I'm current used an "ConvNet" model based on the "ConvNet_CIFAR10_DataAug.cntk" example to classify images into 6 category. The trained error rate is below 1%. However, the error rate is much higher with untrained images. The image size is 128x128x1 and about 10,000 images were used on the training.

The question is what are the methods that people normally try to improve the classification results? I tried using "renet" model and was not able to obtain improvements.

Any input would be greatly appreciated.

Thanks, Terry

Answer 1

This is a classical instance of overfitting. Here are a few things to do:

• Get more data (you could try data augmentation)
• Try adding a drop out layers
• Add regularization (l1/l2) in your learner
• Try early stopping (use a model that has larger training error) and a comparable testing error
• Try reducing the model parameters if you can.