What influences fluctuations in validation accuracy? This is the Tensorboard output of some machine learning experiments I'm doing.

What are the factors that influence the entity of the fluctuations in the validation accuracy? For example: the batch size, the optimizer, regularizers...
Thanks
 A: First of all, does your $x$ axis represent training steps or epochs?
My guess would be epochs (keras default), because of the stability in the training accuracy. If that is not the case, a low batch size would be the prime suspect in fluctuations, because the accuracy would depend on what examples the model sees at each batch. However, that should effect both the training and validation accuracies.
Another parameter that usually effects fluctuations is a high learning rate. The weights change much in each epoch, resulting in the model changing its prediction on many examples. 
Normally this should effect both training and validation sets, but you also seem to be suffering from a bit of overfitting. If this is the case, your model has learned its training set by heart (I can't see this to confirm it by I suspect your training accuracy is close to 1), but struggles a bit on the validation set. This along with a high learning rate would result in the training and validation figures above.
My suggestions to counter this would be:


*

*Decrease the learning rate. Some ideas would be a gradual decrease, a scheduled decrease or a reduction on a plateau of a training metric. I'd recommend the third (can be done easily through a keras callback).

*Regularize the model. This should reduce overfitting and also improve the performance of the model. However, this might increase the fluctuations in the training set as well.
