How To Increase Accuracy in Pokemon Classifier using NN and MLP

#1

Here the accuracy on testing test is 55%.Prateek Bhaiya told this is due to overfitting.But Now how can we increase the accuracy of testing set?

#2

hey @Sid10 ,
while working with neural networks , there are many things you need to think before making your final model for inference.

  1. Check the size of dataset and some sample images too.
    a. If the size of dataset is very small , like around 500 - 800 , then you should first increase your dataset size by applying several augmentations those images as neural networks needs large amount of data to learn.
    b. If your dataset is quite large , more than 1000 records , then too if you want too , you can perform augmentations but there is no such need in this case.
    c. Understanding the dataset is quite important , for example for a particular class , there are some images with white background , some have black background , some look like cartoons , some are bit blurry. Due to these kind of variations in data , the network is bot able to understand them properly and this can be a reason for the model not performing well.

So Now you have got a understanding of how to preprocess the data for our networks.
Now the time comes for creating networks.

While creating Networks you need to think with your perspective that which layer should i use and what number of nodes it should contain. CNN’s works best for retrieving information from images . So you need to make a number of experiments to get the best model.
Then comes to rescue is transfer learning , in which we use a pre-trained model to generate the information from images and using this information , we pass this information into our feed-forward network which work on it and provide us with much better results.

I hope this would help you .
Thank You and Happy Learning. :slightly_smiling_face:.