cross entropy loss is calculated over a two probability distribution, but accuracy is either or kind of calculation. so if my I predict [[0.51, 0.49], [0.49,0.51]] and the real values are [[0,1], [1,0]] I will have a very low cross entropy but accuracy is 0%. but if my predictions are [[0.49, 0.51], [0.01, 0.99]] with the same targets I have 50% accuracy but much higher cross entropy loss because I am not very confident in my right guesses but very confident in my mistakes.