[Review]: Introduction to deep learning

[svenvanderburg]

Lesson Title

Introduction to deep learning

Lesson Repository URL

https://github.com/carpentries-incubator/deep-learning-intro

Lesson Website URL

https://carpentries-incubator.github.io/deep-learning-intro/

Lesson Description

This is a hands-on introduction to the first steps in Deep Learning, intended for researchers who are familiar with (non-deep) Machine Learning.

The use of Deep Learning has seen a sharp increase of popularity and applicability over the last decade. While Deep Learning can be a useful tool for researchers from a wide range of domains, taking the first steps in the world of Deep Learning can be somewhat intimidating. This introduction aims to cover the basics of Deep Learning in a practical and hands-on manner, so that upon completion, you will be able to train your first neural network and understand what next steps to take to improve the model.

We start with explaining the basic concepts of neural networks, and then go through the different steps of a Deep Learning workflow. Learners will learn how to prepare data for deep learning, how to implement a basic Deep Learning model in Python with Keras, how to monitor and troubleshoot the training process and how to implement different layer types such as convolutional layers.

Author Usernames

@dsmits @psteinb @cpranav93 @colinsauze @CunliangGeng

Zenodo DOI

10.5281/zenodo.8308392

Differences From Existing Lessons

No response

Confirmation of Lesson Requirements

• is the original work of the author(s), or that any content derived from another source is reused with permission and appropriate attribution
• aligns with The Carpentries Code of Conduct
• is published under a CC-BY or CC0 license
• uses The Carpentries lesson template or Carpentries Workbench without significant customisation/adaptation.

JOSE Submission Requirements

• the lesson repository includes paper.md and paper.bib files as described in the JOSE submission guide for learning modules

Potential Reviewers

No response

[svenvanderburg]

We're still running a final round of comments for the paper (see carpentries-incubator/deep-learning-intro#364), the plan is to submit the paper on the 15th of September through: https://openjournals.readthedocs.io/en/jose/submitting.html#submitting-your-paper . I'm not sure how that relates to this review, are the 2 independent or should we wait for this review process to finish before submitting to JOSE?

[tobyhodges]

Great to see this submission, @svenvanderburg. As a listed contributor, I have a conflict of interest acting as editor for this one. I am going to find another community member who can fulfill the role for this review, and will post back here when it's ready.

To answer your question about review order: we have been asking lesson developers to submit the lesson for review here, before the JOSE review. See #11 and the related review in JOSE for an example.

[svenvanderburg]

@tobyhodges any update on finding an editor?

[svenvanderburg]

@tobyhodges ? 😇

[tobyhodges]

I have reached out to a potential guest editor for this review and am waiting for confirmation. Hoping to be able to follow up very soon!

[svenvanderburg]

I have reached out to a potential guest editor for this review and am waiting for confirmation. Hoping to be able to follow up very soon!

Perfect, thank you for the update 🙏

[tobyhodges]

Good news: @brownsarahm has kindly agreed to act as Guest Editor for this review. I am extremely grateful to her for being willing to take this on.

[brownsarahm]

I'll work on this in small bits, but this way it's all in one place and the authors could work on the (very minor) accessibility issues and one small note on setup that I have checked so far.

Editor Checklist - Intro to Deep Learning

Accessibility

• All figures are also described in image alternative text or elsewhere in the lesson body.

• The lesson uses appropriate heading levels:

  • h2 is used for sections within a page.
  • no “jumps” are present between heading levels e.g. h2->h4.
  • no page contains more than one h1 element i.e. none of the source files include first-level headings.

• The contrast ratio of text in all figures is at least 4.5:1.

• the check boxes in the prereqs render poorly in the workbench + are an accessibility flag bc they're unlabeled "form elements" these should be removed

• listed browser support is not matched to jupyter browser support notably, Windows' current default browser is a chromium browser and therefore supported (and on general web compatibility, better than safari by a nontrivial margin) this should be updated to include more windows users without having to install

• most table headers are empty, but this might not be a problem

• in ep 2 there is a link to the sklearn docs with the text "here" this is rated "suspcicous" in accessibility terms, recommend "explained in the scikit-learn docs" instead of explained "here"

• minor syntax issue (: instead of =) causing "missing alt text" on an image in :

  • ep 1
  • ep 4 link to source

Content

• The lesson teaches data and/or computational skills that could promote efficient, open, and reproducible research.
• All exercises have solutions.
• Opportunities for formative assessments are included and distributed throughout the lesson sufficiently to track learner progress. (We aim for at least one formative assessment every 10-15 minutes.)
• Any data sets used in the lesson are published under a permissive open license i.e. CC0 or equivalent.

Datasets and licenses

• penguin in CC0
• weather is CC-BY
• CIFAR10 license is unclear/not easy to find/ may not exist, but it is public data

other content notes:

• deep learning workflow exercise in ep 1 solution is not a solution; there can be many answers here but maybe some details of what to check for in an answer would help
• "just" used dropna in clean missing values is not Carpentries language and can be read as dismissive. "ruin the training data" is also not very precise about the impact and does not acknowledge that there are different ways to handle missing data for reasons that can impact analyses. In line with teachign best practices, a slightly broader coverage/more intentional choice of using dropna, which is fine to do, would improve the lesson here
• ex reflecting on our resutls is oddly formatted and has an exercise in the solution
• solution to varying the dropout rate is a term that was introduced in a different form in a parenthetical in the previous episode, but not defined as it is used.

Design

• Learning objectives are defined for the lesson and every episode.
• The target audience of the lesson is identified specifically and in sufficient detail.

to fix:

• question, objectives, and keypoints all display with "" around them

Repository

The lesson repository includes:

• a CC-BY or CC0 license.

• a CODE_OF_CONDUCT.md file that links to The Carpentries Code of Conduct.

• a list of lesson maintainers.

• tabs to display Issues and Pull Requests for the project.

• replace this with any further comments relating to the lesson repository.

Structure

• Estimated times are included in every episode for teaching and completing exercises.
• Episodes lengths are appropriate for the management of cognitive load throughout the lesson.

comment:

• all episodes are 55min + episode 3 is over 3 hours long?

Supporting information

The lesson includes:

• a list of required prior skills and/or knowledge.
• setup and installation instructions.
• a glossary of key terms or links out to definitions in an external glossary e.g. Glosario.

other setup note:

• it says "open a terminal" but that is insufficient information for a Windows user (they have multiple terminals and to launch jupyter only the "anaconda prompt" by default will work; CMD or powershell typically will not)

General

• replace this with any other comments that do not fit into any of the previous sections.

[svenvanderburg]

@brownsarahm any update on the progress?

[svenvanderburg]

@brownsarahm any update? Can you give us an indication when we can expect this to be done?

[brownsarahm]

Hi! Sorry, a bunch of unexpected things happened last fall, and then when you sent the first check-in I was off of work for the holidays. And the second came while I was in a deadline crunch,

This is now back in my active queue. I should finish the pre-reviewer stuff within a week and I'm looking for reviewers starting now.

[svenvanderburg]

Hi @brownsarahm. Cool, thanks for the update 🤗

[brownsarahm]

editorial checks are done and @tobyhodges and I are working on finding reviewers next.

The comment above has a few things for you all to look at now, but thanks for resolving all of the previously identified ones already!

[tobyhodges]

Thanks @brownsarahm. I have suggested a few reviewers to Sarah but if any of my fellow authors can also suggest anyone they think would be suitable, I am sure it would be helpful. (Please do not tag anyone here by their GitHub handle.)

[brownsarahm]

@svenvanderburg Do you have any updates in response to my final editorial check?

In particular, do you have responses to the concerns about:

• episode lengths
• dataset licenses

and ideally before we assign reviewers, it would be nice to resolve, but these are minor:

• " rendering on episode metadata
• typos breaking alt-text

[svenvanderburg]

Hey @brownsarahm.

Sorry, I totally missed that it was final! Thanks for extra pinging me @brownsarahm :)

Some answers:

Episode lengths

Indeed episode 3 takes a bit longer than the other ones, but not twice longer. In fact, the timing for the other episode is too optimistic: Here is a PR with more realistic timing. In comparison to other lessons the episodes are a bit longer, this is because we want to finish the full deep learning workflow in each episode. When teaching, this is not really a problem though, the workshop is actually pretty balanced in terms of cognitive load because in every episode we go through this deep learning workflow once, and conceptually it makes sense to have the cuts between episodes at these points. What do you think? We could maybe write this explanation in the introductory instructor notes?

Dataset licenses

If I understand correctly, the only problem is with the CIFAR-10 dataset. It is so widely used, but I never realized it doesn't actually have a license judging from the official website.

I found a paper that extracts the statement about citation as their license:
Running example. In this paper, we download the CIFAR-10 dataset from its official website. Also on the CIFAR-10 website, we find the following request from the dataset creators: “Please cite (Krizhevsky et al., 2009) if you intend to use this dataset," alongside a link to the paper. We extract this as the dataset’s license.

It is actually a crawled dataset, and in the paper I don't read anything about the crawled images being under open-source license....

Anyway: do you think it is a problem that we use this dataset? It is such a central dataset in the field and so widely used. We would have to change the entire episode if we use a different dataset. We can write a comment about in the instructor notes.

Small issues

I resolved the two small issues you referred to, will be reviewed soon by one of my colleagues. We will soon pick up any other remaining issues.

• Small fixes from Carpentries Lab review carpentries-incubator/deep-learning-intro#444

Would be good to enter the next phase of reviewing. Let us know what we have to do to help this progress.

[brownsarahm]

Timing

Thanks for the explanation. I think at one level your strategy for putting breaks in the content makes sense. I am not certain if your claim about cognitive load is true, but also not certain that it is false.

However 3.5 hours without a break is a long time and most instructors will not want to give a break in the middle of an episode.

Maybe an instructor note reminding about breaks? (as context, I'm a maintainer on instructor training and we get a lot of complaints about not enough breaks there and we have them ~every 90 minutes).

Dataset License

In my understanding of carpentries policy a permissive license is required. Since the dataset was crawled, it likely has zero consent to be using the images from the original owners of the images. I think it probably does not have the same risks as imagenet, but should be checked. On the other hand, this is clearly, to me, an intended use by the people who curated the images into a dataset, despite them not putting a license on it and them possibly not having appropriate rights to the images either.

Whether it is okay or not is going to be up to Carpentries policy about the situation where there is no license. @tobyhodges can you help navigate that or let us know who else in the carpentries should be looped in?

[tobyhodges]

Thanks for tagging me @brownsarahm. I need to do a bit more reading and thinking about this, and will come back soon with a full response.

[tobyhodges]

Thanks for your patience while I took some time to read through the relevant pages and documents, and to reflect on the most appropriate course of action. I am sorry to say that I think we should replace the dataset in the lesson.

The lack of a license file in the dataset is somewhat problematic, even though the authors clearly intend for the dataset to be re-used and usage in the lesson is within the terms stated on their website. But my biggest concerns are with the unethical way in which the data was "collected." Images were scraped and modified for the dataset without any attempt at seeking permission from the copyright owners or giving them attribution, which feels unethical to me regardless of any arguments over its legality. (I am not a lawyer but it seems like the use may fall under "fair dealing" in Canadian copyright law, where the researchers who published the dataset are based.)

In Collaborative Lesson Development Training, alongside considerations of licensing, size, and complexity, we ask lesson developers to consider the ethics of the example datasets they include in their lessons. I would like to apply the same standard to lesson reviews in The Carpentries Lab.

I acknowledge that replacing the dataset will require significant new work on the part of the authors, and perhaps I should have noticed sooner and avoided some of this inconvenience. @svenvanderburg for my part, I would like to devote some time in the coming weeks to try to make the necessary changes (as I am already one of the authors). I hope I will be able to propose some alternative datasets soon, and of course it would help to have input from others with more DL experience than I have. However, please also be aware that you can withdraw the lesson from review here if you prefer.

Finally, many thanks to @brownsarahm for catching this and looping me into the discussion.

[svenvanderburg]

Timing

Ha, no 3,5 hour teaching without breaks is absurd! At the Netherlands eScience Center we usually teach in a schedule like [for this recently taught workshop](this https://esciencecenter-digital-skills.github.io/2024-02-05-ds-dl-intro/#schedule). Never more than 90 minutes of teaching! And I think beta pilots copied that schedule in rough lines. It doesn't matter that it is in the middle of an episode.

In addition, we swap instructors halfway the episodes which makes teaching load lighter as well.

See carpentries-incubator/deep-learning-intro#446 for addressing this, do you agree @brownsarahm ? And thanks for bringing this up, this is a great outcome of the review. Since we always use the same schedule no matter what lesson material we use we had a blind spot here.

License

Thanks @tobyhodges for digging into the CIFAR10 license. I agree we should change it, indeed it goes against everything the Carpentries stands for...

So, the remaining issues to fix before the review are:

• Deal with relatively long episodes in instructor notes carpentries-incubator/deep-learning-intro#446
• CIFAR10 not having a license carpentries-incubator/deep-learning-intro#445

(and some more small comments from Sarah that we will definitely pick up the coming period but are not essential to do before the review)

Can you confirm this @brownsarahm ?

[brownsarahm]

Yes, this is correct, these two issues would get it to a point where it is ready for review.

[svenvanderburg]

@likeajumprope can you update us on the progress of your review? (Where no progress is also a valid update 😉 )

[brownsarahm]

@likeajumprope Do you have progress on this? Even an incomplete review helps the authors improve the lesson and then you can update after.

[likeajumprope]

Reviewer Checklist

Accessibility

• The alternative text of all figures is accurate and sufficiently detailed.

  • Large and/or complex figures may not be described completely in the alt text of the image and instead be described elsewhere in the main body of the episode.

• The lesson content does not make extensive use of colloquialisms, region- or culture-specific references, or idioms.

• The lesson content does not make extensive use of contractions (“can’t” instead of “cannot”, “we’ve” instead of “we have”, etc).

• replace this with any further comments relating to the accessibility of the lesson.

Content

• The lesson content:

  • conforms to The Carpentries Code of Conduct.
  • meets the objectives defined by the authors.
  • is appropriate for the target audience identified for the lesson.
  • is accurate.
  • is descriptive and easy to understand.
  • is appropriately structured to manage cognitive load.
  • does not use dismissive language.

• Tools used in the lesson are open source or, where tools used are closed source/proprietary, there is a good reason for this e.g. no open source alternatives are available or widely-used in the lesson domain.

• Any example data sets used in the lesson are accessible, well-described, available under a CC0 license, and representative of data typically encountered in the domain.

• The lesson does not make use of superfluous data sets, e.g. increasing cognitive load for learners by introducing a new data set instead of reusing another that is already present in the lesson.

• The example tasks and narrative of the lesson are appropriate and realistic.

• The solutions to all exercises are accurate and sufficiently explained.

• The lesson includes exercises in a variety of formats.

• Exercise tasks and formats are appropriate for the expected experience level of the target audience.

• All lesson and episode objectives are assessed by exercises or another opportunity for formative assessment.

• Exercises are designed with diagnostic power.

• replace this with any further comments relating to the lesson content.

Design

• Learning objectives for the lesson and its episodes are clear, descriptive, and measurable. They focus on the skills being taught and not the functions/tools e.g. “filter the rows of a data frame based on the contents of one or more columns,” rather than “use the filter function on a data frame.”

• The target audience identified for the lesson is specific and realistic.

• replace this with any further comments relating to the design of the lesson.

Supporting information

• The list of required prior skills and/or knowledge is complete and accurate.

• The setup and installation instructions are complete, accurate, and easy to follow.

• No key terms are missing from the lesson glossary or are not linked to definitions in an external glossary e.g. Glosario.

• replace this with any further comments relating to supporting information for the lesson.

General

• replace this with any other comments that do not fit into any of the previous sections.

[likeajumprope]

Almost done - in general the lesson is super nice and I think I can check most boxes above. Will fill in the general section with my notes in the next few days

[svenvanderburg]

Great @likeajumprope 🎉 Looking forward to your notes 😺

[svenvanderburg]

@likeajumprope any update on this? 😽

[svenvanderburg]

@likeajumprope we have a sprint upcoming next week Monday 13 & Tuesday 14th of January. It would be awesome if you could share your feedback with us so we can improve the lesson.

[likeajumprope]

@likeajumprope we have a sprint upcoming next week Monday 13 & Tuesday 14th of January. It would be awesome if you could share your feedback with us so we can improve the lesson.

Hi @svenvanderburg I think I can do that - I have been working on it over the Christmas Holiday and I am almost ready. Sorry for the delay and thanks for being so patient - I know I am stretching it a bit

[svenvanderburg]

Hi @likeajumprope great to hear from you, happy new year ⭐️ If you could do that, that would be awesome!

[likeajumprope]

Hi @svenvanderburg, please see below my detailed feedback and a print out of the Jupyter Book.
Overall, fantastic tutorial! See below for the few comments and the sections that gave me errors.

Carpentries_DeepLearnJupyter.pdf
DeepLearning_feedback.pdf

[svenvanderburg]

Great @likeajumprope, thanks a lot, especially given this was quite a ride! 🎉🙏 We will work on your comments in our lesson development sprint today and tomorrow.

[likeajumprope]

Great @likeajumprope, thanks a lot, especially given this was quite a ride! 🎉🙏 We will work on your comments in our lesson development sprint today and tomorrow.

Thanks for sticking with me ;)

[svenvanderburg]

@likeajumprope thank you for the useful comments! We resolved almost all your comments in carpentries-incubator/deep-learning-intro#523. For 2 comments we need a few clarifications:

Sketch of a neuron

'Introduction of the neuron: It might be useful to actually show the sketch of a neuron here – not everyone might know what a neuron looks like.' Referring to: https://carpentries-incubator.github.io/deep-learning-intro/1-introduction.html#neural-networks

Do you mean to show a picture of an actual neuron? I think that might actually confuse learners because you go into the discussion of which properties of an actual neuron than correspond to that of an artificial neuron.

Mean Squared Error formula

You mention 'Compute the Mean Squared Error: the hat is over the bracket instead of the y, in the MSE equation', regarding https://carpentries-incubator.github.io/deep-learning-intro/1-introduction.html#compute-the-mean-squared-error . Can you let is know in which browser you see this. In chrome I get the hat above the y:

[carschno]

As promised, here's the remaining comments. I'll post the reviewer checklist after this along with overall comments/summary. Again, i've very happy with the lesson and most of my comments are aimed at improving the clarity / further reducing the cognitive-load of a fairly heavy topic! (the lesson does a very good job already).

I have gathered these issues in the tasklist of this issue

A few specific replies @mike-ivs :

in the "BatchNorm" section we redefine create_nn again. Is it worth giving this a different name for clarity/to avoid overriding the earlier model?

I think it is worth re-creating the function here because it introduces new functionality every time. This way, we provide the full context so that learners know exactly where to put it.

I realise it's in a potentially "watch only" section, but should we move this to pre-lesson installation instructions? Is there a good enough reason install part-way through? It's potentially an optional section, so I don't feel too strongly about it either way.

I think the reason is that it is not required for anything in the lesson and we don't want to install unnecessary packages.

A diagram would help clear this up by explaining where we add our own architecture wrapper around the existing model.

I don't necessarily agree. I find the diagram that you have linked potentially confusing rather than enlightening. I am sure there could be a good diagram, but I have instead tried elaborate in the text a bit.

[mike-ivs]

@carschno Thanks for the reply. I'm happy with all those answers :)

Specifically: Looking at the transfer-learning diagram I linked 6 months ago I would now agree with you... it's more confusing that helpful! The changes you made help clear things up. Great work on the lesson.

[brownsarahm]

I am using chrome (Version 131.0.6778.86 ) and see the MSE equation rendered poorly:

I tried resizing my browser window and checked light/dark mode toggle and it stayed bad. Since the source is correct, let's ask @tobyhodges about what might cause the rendering error and call this a workbench issue, not a lesson issue?

[carschno]

Thanks for checking that @brownsarahm, very sharp eyes! :)

I see the same on Chrome on a Mac Version 131.0.6778.265 (Official Build) (arm64).
Firefox displays the equation correctly, see screenshot):

I agree this looks like a workbench and/or browser issue.

[tobyhodges]

Indeed. This is an issue with varnish, the component of the Workbench that applies styling etc to the lesson website. Specifically, it is something to do with how the MathJax JavaScript library is rendering the equation in Chrome (broken on when viewed in that browser on my system too). I am investigating further and will open an issue and/or PR on https://github.com/carpentries/varnish in due course. There is nothing to be done on the DL lesson repository to resolve this in the meantime.

[brownsarahm]

that it worked in firefox made me think it's not a workbench issue, so i did a quick search and found a solution that attributes it to macos but is a quick change to make:

1. Right-click on some of the misrendered MathJax.
2. Click on "Math Settings".
3. Click on "Math Renderer".
4. Click on "Common HTML".

Maybe it is worth putting a tip box somewhere in the setup or instructor notes with this solution since readers of the lesson can fix it for themselves?

[carschno]

Maybe it is worth putting a tip box somewhere in the setup or instructor notes with this solution since readers of the lesson can fix it for themselves?

I have added a PR: carpentries-incubator/deep-learning-intro#539

[tobyhodges]

Great catch @brownsarahm, thanks 🙌

[carschno]

With carpentries-incubator/deep-learning-intro#530, the final pending PR to address the review comments has now been merged.
Are we ready to finalize the review then @brownsarahm?

[brownsarahm]

I had forgotten to update this to step 4 when the second review came in but we have now jumped to step 5! 🎉

@likeajumprope and @mike-ivs please take some time to look through the authors' responses above and the updated version of the lesson. When you are ready, post back here to ask follow-up questions, let the authors know about any additional changes you would like them to make, or recommend that we accept the lesson to The Carpentries Lab. You may find it helpful to refer to the Reviewer checklist again if you would like to use that as a guide.

We are so close team!