TF1 C2 W2 quiz question needs more info

The quiz includes this question:

When training with augmentation, you noticed that the training is a little slower. Why?

From trial and error, I believe this question is unanswerable without additional information. I propose editing the question to this:

When training with augmentation leading to N total training examples, you noticed that the training is a little slower than it is when you train with N training examples which include no augmentation. Why?

The staff have been notified about your concern.

Given that the quiz question belongs to week 2, would you please change the topic title to W2 from W1? Thanks.

1 Like

Done. Thanks for the heads up!

The staff have fixed the quiz question.

1 Like

Hi Balaji, I just reloaded the quiz to see the change implemented and unfortunately the old phrasing of the question is still there. Please advise.

I just opened the quiz and noticed the updated question. You might want to clear your cookies and try again.

Hi Balaji, Thanks for this. The question appears to be updated but not in the way that would make the “correct” answer correct. My understanding is that the (answer deleted) option is correct if the augmentation process actually creates more image data for training, and all of the lecture notes so far clearly imply that it does. For example, without augmentation maybe we have 1000 images and after maybe we have 20000. In order for the (answer deleted) option to be correct, the question would have to clarify that the net number of actual examples used in training before and after augmentation is in fact the same. In other words if we start with 1000 images and after augmentation somehow still have 1000 images (I am not sure why we would do this) then the wording I proposed makes that clarification.

However, if I misunderstand what augmentation is doing, and that it is not creating many times more training images, then can you please clarify what it is doing?

The number of images per epoch is constant whether or not augmentation happens. The question asks about the impact of augmentation on the training time. So, the staff provided answer is correct.

Please remove quiz answers from your latest reply.

Hi Balaji, Thanks for your commet. I just removed the details from my earlier comment that I believe you were referring to.

Regarding the staff answer, I did some research to sanity check my initial claim. Here are the details:

  • Data augmentation creates more data. This is stated a few times in the lectures, and is confirmed by a web search. The lectures also clarify that the augmentation does not impact the original images, which is a nice benefit. However, the fact that keras provides this benefit does not mean that additional images are not used in training. They are.
  • Training with more data is slower. This is common sense (more computation takes longer) and is also confirmed by a web search. Using more training examples leads to more batches, larger batches, or both.
  • Therefore, data augmentation, which creates more data, leads to slower training.

Given the above, there are two or potentially three correct answers listed to the quiz question. (It depends on what is meant by “data” being “bigger.”) Can course staff please update the question to have only one correct answer? I provided one way to do this, but it’s just an example. Hope this helps.

For anyone still following, I pasted the question into ChatGPT, whose answer is in agreement with what I posted above, which is also in agreement with what I found in the lecture notes and what I found via websearch. The “correct” answer offered in the quiz is at odds with all of these. Course staff, are you able to see these threads?

Data augmentation involves creating new synthetic training examples by applying various transformations, such as rotations, translations, and flips, to the existing training data. These transformations effectively increase the size of the training set, which can help improve the performance of the model by reducing overfitting and improving its ability to generalize to new data.

However, because data augmentation effectively increases the size of the training set, training a neural network with augmented data takes longer than training without it. The model needs to learn from more examples, and each epoch of training will take longer as there are more examples to process.

Additionally, the augmentation process itself can also add some computational overhead to the training process. The additional time required to apply the transformations to the training data and feed it into the neural network can contribute to longer training times.

Despite the longer training times, data augmentation can be a powerful technique for improving the performance of neural networks, particularly in cases where the available training data is limited.

Hi Aaron! Thank you for bringing this up, and sorry for the late resolution. These are good points indeed and show the ambiguity in the question. We revised it to mention the training epoch time, instead of just “training”. Hopefully, this clarifies it for future learners. Thanks again!

Thanks Chris. Since I’m no longer enrolled in the course, I can’t browse to the question to see if the update addresses the noted issue. Can you post the update here, and optionally DM me the answer course staff has marked as correct?