So, I have the following charts from my experience.Can any one explain why accuracy is decreasing while the loss in train and validation is decreasing? The point is that i can't early stop too in the maximum accuracy point, because of the decreasing validation loss in that point.
2 Answers
First of all, you can perform early stopping on any metrics. If you are using Keras you can do something like
es = EarlyStopping(monitor='val_accuracy', mode='max', min_delta=1)
which is taken from this post. If you want to do it by hand it's enough to e.g. save the weights each time you improve your accuracy.
Additionally, if accuracy is your ultimate goal, maybe you have the wrong loss (since as you say it keeps decreasing). In such case you have to find a surrogate loss (see this post) of the accuracy for your specific problem. This can be very hard to do, but there are some papers that explore this idea.
Can any one explain why accuracy is decreasing while the loss in train and validation is decreasing?
This can happen whenever accuracy is not monotonic function of the loss. In that case it may be possible to improve the loss (which is probably the main objective that your algorithm tries to achieve) while the accuracy does not improve or even decreases.