Retraining approches

I have an overfitting problem in my model (training accuracy 100%, testing accuracy 65%), I should retrain the model with only new samples or I should combine the new samples with the old samples and retrain the model from scratch.

However, I need the old samples not combined with the new samples because the old samples are easily recognizable samples and new samples are difficult. In brief, I want to apply curriculum learning but I encountered the problem of overfitting.

1 Like

On your initial data you may retrain with augmentation to avoid overfitting.

Do share your performance graphs along with no. of samples in train and test sets to check data distribution

1 Like

No. samples are 1500 taken from MNIST.

In fact, I used Curriculum learning to sort my data, and then as the curriculum learning protocol, you should start the training with easy samples which are 1500 samples (in my case) and then I stuck with how to train the neural network. I trained the network with these easy samples 20 iterations and then I want to move to the harder ones but I was surprised with my model overfitting the data. It is hard to continue training with overfitted model. Finally, I know that there is a pace function that tells you the pace of training but I did not understand very well the output of this function and how and when I should use.


1 Like

Explore albumentations augmentation library.

Category icons made by Freepik from