C4W2 Quiz has two errors, fixes proposed

Issue 1

True / False question:
Parameters trained for one computer vision task can’t be used directly in another task. In most cases, we must change the softmax layer, or the last layers of the model and re-train for the new task.

Comment: This question as currently worded is false, not true. Specifically, the first sentence of the prompt is false. In the lecture for the “Transfer Learning” slide, Prof Ng explicitly describes, on the first third of that slide, that we can “freeze” parameters for the first several layers and training only the last layer.

Proposed fix: if you want the answer to this question to be true, change the prompt wording to: “The full set of parameters and layers trained for one computer vision task can’t be used directly (without change) in another task. In most cases, we must change the softmax layer, or the last layers of the model and re-train for the new task.”


Issue 2

Prompt: “Which of the following do you typically see in a ConvNet? (Check all that apply.)”

Comment: the many ConvNet examples shown in the lecture notes are consistent: alternating conv and pool layers followed by FC layers. However, when I answered the question this way (selecting only boxes corresponding to such networks), I was told I was wrong. The proposed fix depends on whether the lecture notes are showing us typical exampes or not. Thus I have proposed two fixes.

Proposed Fix 1: if the lecture notes’ examples are representative, then do not require learners check boxes other than those indicating alternating conv and pool layers followed by FC layers. In other words, the quiz is wrong and would need to be fixed by eliminating wrong choices from the “correct” choices.

Proposed Fix 2: if there are different types of ConvNets that are not alternating conv and pool layers followed by FC layers, redo the lectures and lecture notes, so that when this comes up on the quiz, learners will answer correctly.

Hey @am003e,
Apologies for the delayed response. It would be great if you can help understand these issues better, so that we can take the required steps.

Issue 1

In my opinion, the word “directly” which has been included in the statement, essentially implies “without changing at all”, and I believe that is true. Had this word been omitted, then it would be more relevant to do the proposed fix. What are your thoughts regarding this?

Issue 2

Can you please let me know the question, and the options that you received in your quiz, and also highlight the options which require modifications in your opinion? I am a little confused as to which of the variations of this question you received in your quiz.

Cheers,
Elemento

Regarding issue 1, the conflict is over the word “parameters.” In standard English, using a qualified collective noun in a context like this generally implies exhaustive representation of that noun. When we say “Parameters trained for one computer vision task can’t be used directly in another task” it means that each of the individual parameters we’re talking about (those trained for one computer vision task) can’t be used for another. However this is false because some of the parameters - in fact, most of them - can be used unmodified. The proposed wording fixes the issue. To be even more explicit:

  • Parameters trained for one computer vision task can’t be used directly in another task. ← FALSE, because some of the parameters can be, but the wording implies we’re talking about all of them.
  • Some of the parameters trained for one computer vision task can’t be used directly in another task. ← TRUE, because we qualified it.
  • The full set of parameters and layers trained for one computer vision task can’t be used directly (without change) in another task. ← TRUE, because we clarified that it’s the full entirety of the set that can’t be used unmodified, even if some of the individual parameters in the set could be used unmodified.
  • Some of the parameters and layers trained for one computer vision task can’t be used directly (without change) in another task. ← TRUE, because we clarified here that it’s only some of the parameters.
    Helpful?

Regarding Issue 2, I unfortunately don’t have course access anymore as I completed it long ago and my subscription has expired. You can DM me the question options and I’ll tell you which are at odds with the lectures.

Hey @am003e,
I get your point regarding the issue 1. I believe it is delving too deep into the waters of the English language. But still let me create an issue for this, and if the team deems fit, they will change it accordingly.

Cheers,
Elemento

Thanks Elemento. Given your reply, can you please post here if this winds up getting resolved? In the meantime, I’ll consider this one open; I’m separately tracking how many quiz errors remain since this impacts my course review.

Hey @am003e,
Sure, I will let you know the updates regarding this. Once again, thanks a lot for your detailed feedback :nerd_face:

Cheers,
Elemento

Hey @am003e,
Issue 1 has been resolved. In this, the team has modified the sentence from “Parameters trained for one computer vision task …” to “Models trained for one computer vision task …”. It might take some time to reflect in your account. And as for issue 2, I hope that was resolved too, from our discussion in the private thread. Once again, thanks a lot for the feedback :nerd_face:

Cheers,
Elemento

Hey @am003e,
Issue 2 (Question 1) has been resolved. As discussed in our DM, in variation 1, there are no discrepancies, and in variation 2, the following change has been made:

Use of FC layers after flattening the volume to generate output classes

Cheers,
Elemento

Use of FC layers after flattening the volume to generate output classes

This is close to correct. In order for this interpretation to be correct, it must be parsed as follows:

(Use of FC layers after flattening the volume) [whose output is then used] to generate output classes

However, the more idiomatic reading of this is as follows

Use of FC layers after (flattening the volume to generate output classes)

which is incorrect. I proposed this update in my email:

Use of FC layers after flattening the volume with the last FC layer mapping to output classes.

but it appears that solution has been declined. Learners: if you see this come up on a quiz, please be informed that they wish for you to assume the first interpretation. Of course, it’s probably quicker to just fail this item on the quiz and then try again. :slight_smile: