While training a neural network with mini-batches for a classification task, the training accuracy fluctuates wildly between iterations. What is the most likely cause?

A high learning rate causes large updates to the model's weights during training. This can lead to the optimization algorithm overshooting the optimal solution, causing the loss function and accuracy to fluctuate erratically instead of converging smoothly. The model struggles to find a stable minimum. While an imbalanced dataset can affect overall accuracy and lead to biased predictions, it typically doesn't cause wild fluctuations between iterations in the way a high learning rate does. Disabling shuffling between epochs means the model sees data in the same order each epoch, potentially leading to slower convergence or getting stuck in local minima, but not necessarily wild per-iteration fluctuations. A mini-batch size that is too large can lead to slower training and potentially getting stuck in local minima, but it generally results in smoother, albeit slower, updates, not wild fluctuations.