Problems with the existing approach
In this section, we will list all the points that create problems. We should try to improve them. The following are things that I feel we can improve upon:
If you find out that class sampling is not proper in your case, then you can adopt the sampling methods
We can add more layers to our neural network
We can try different gradient descent techniques.
In this approach, training takes a lot of time that means training is computationally expensive. When we trained the model, we used GPUs even though GPU training takes a long time. We can use multiple GPUs, but that is expensive, and a cloud instance with multiple GPUs is not affordable. So, if we can use transfer learning in this application, or use the pre-trained model, then we will achieve better results.
