Consider the following scenario. Given a dataset with 5000 instances and 20 attributes(including the class label), we used MLP with various hyperparameters, and also cross validation, but still the accuracy is low(during training and testing), we need to try to improve the accuracy, the class labels are balanced in the datasets, and we do not have anymore data to add to our system, we could not use dropout because the training accuracy was low. Suggest an approach/s with clear steps to improve the accuracy (think out of the box, you may think to use hybrid approach)