Two questions about observed training accuracy during training of a convolutional neural network such as cats-vs-dogs or horses-vs-humans.
Oftentimes the accuracy at the start of an epoch is significantly higher than that at the end of the previous one. Why?
After the initially reported accuracy at the start of the epoch, the accuracy typically goes down. Why?
Is it just a coincidence?
Okay so what happens within an epoch is not completely predictable…It depends on the data the model gets,its predictions and various other parameters too
It usually happens what you stated starting with very high and then decreasing a bit to level off at a point
You would notice this trend but as of to why this occurs is not completely predictable or even of it is ,it is way too complex I think
As of the increase in accuracy between 2 epochs yes that has to increase as the model learns more and more with the epochs