Exam AWS Certified Machine Learning - Specialty topic 1 question 153 discussion

D - over fitting problem.

The specialist should consider using some form of regularization to fix this issue. Regularization techniques such as dropout or L2 regularization can help prevent overfitting, which can occur when the model performs well on the training data but poorly on the validation data. Option A, a longer training time, might not necessarily fix the issue and could lead to overfitting if the model is already performing well on the training data. Option B, making the network larger, could also lead to overfitting and may not be necessary if the current network architecture is sufficient to perform the classification task. Option C, using a different optimizer, might not necessarily fix the issue and could lead to slower convergence or worse performance. Therefore, option D, using some form of regularization, is the most appropriate solution to consider in this situation.

some form of regularization

I wouldn't go with D since it doesn't seem an overfitting problem considering training accuracy is not so high. So the main problem here is to get an higher accuracy even on training set. I would go with A or B

A - IMO it's an underfitting problem, as training accuracy is not better than baseline error (human accuracy). Would consider B as well, but it may actually decrease accuracy.

typical overfitting problem

typical overfitting problem

C - It is not a overfitting problem as the training accuracy stands at 90%, which is at same level of human performance. That means the algorithm used is not optimized for this problem. So, some other algorithm should applied for this problem.

I'd go A. Regularization could not guarantee higher validation accuracy.

I believe answer is B , because clearly it is a overfiting problem , if we reduce complexity the accurate will reduce close to 80% ... But human works can reach up to 90% .