Constant training and validation accuracy while training CNN

Question

I am running CNN on a software defect dataset which I have converted into images using Deep Insight Library. Training the images on the CNN, after every epoch I am witnessing constant training and validation accuracy. I have not used any regularization technique like Dropout, Batch Norm etc. Besides constant training and test accuracies, you all can see that there is a high bias and high variance associated with the model. I will be happy if you can suggest steps that can help me in not getting these constant values and improve my model accuracy. enter image description here

Answer 1

It seems you have fallen into a local minimum. This is fine but may explain that you have gotten to the limits of the capabilities of the current implementation there are a couple of ways to improve the accuracies.

1. Increase dataset size: more data results in beter generalizations usually thus increased accuracies. Datasets should be in the order of 100s to 1000s of items, especially for CNNs. A useful technique is using image augmentation .
2. Increase model complexity: perhaps you have enough data but the model isn't complex enough to solve the problem. So add more depth to the model may help.