Anonymous review 2

[goodfeli]

This is an anonymous review that I am sharing from a peer reviewer. They sent it to me as an e-mail with formatted text. Since I don't know of a way to copy-paste a formatted e-mail into Markdown, I'm just sharing it as an RTF document:

https://drive.google.com/a/google.com/file/d/0Bz8CQw2wxLVwUEF5TTY3SjRoUkU/view?usp=sharing

[colah]

Transcription of review from the RTF document, to preserve record of review in GitHub issues.

---

Thoughts on first read-through:

• The first set of images are really crisp!
  • Clarify what dataset this network was trained on? Are all of these
    examples from the same network? I’m assuming this is a classification
    network, could be nice to make this concrete
• Intro paragraph
  • “state-of-the-art visualizations”, by what metric?
• Optimization Objectives
  • Graphics are great
  • “What do we want examples of?” Good overview that highlights all
    the possibilities of feature visualization. Would be good to state which
    objective(s) the paper is focusing on.
• Why visualize by optimization
  • The images of dataset vs. optimization work well
• Diversity
  • Interesting that one neuron has these clusters of images at min/max
    activation (ie flowers and clocks). Could make it clear what “facets” means
    -- the kinds of images that result in different strength activations of the
    specific neuron? Or is it the kinds of images that result in maximum
    activation?
• Achieving Diversity with Optimization
  • Graphics: make it more clear that rightmost images = w/o diversity,
    and center images are with diversity
  • What diversity term is used in the examples?
• Interaction between Neurons
  • The combination of two neurons is very cool
  • Last paragraph is a good description of the current challenges
• The Enemy of Feature Visualization
  • A bit unclear on why optimization worked in above paragraphs, but
    didn’t over here?
  • Nice, clear overview on regularization options
• Preconditioning and Parameterization
  • Paragraph on transforming gradient is great, makes sense
• Conclusion
  • Possibly include a section on future work in feature visualization?
  • What high-level problems need to be solved?

High-level thoughts

• Paper contributes:
  • A method to get diverse feature visualizations via optimization
  • If the feature visualization is a sort of generative network, simple
    arithmetic or interpolations can be done in feature space
  • An overview of prior methods
• The graphics and visualizations are great

[ludwigschubert]

Thank you for your high-quality feedback! We went through every bullet point and have made numerous changes to the article based upon the review you provided. These can collectively be found in the pull request #7.

We are especially grateful both for the insightful critique as well as the hints about sections that may be hard to understand—we do not just want to be factually correct, but also approachable.

Clarify what dataset this network was trained on? Are all of these examples from the same network? I’m assuming this is a classification network, could be nice to make this concrete

We use GoogLeNet trained on ImageNet.

We added additional captioning on the hero diagram mentioning both the model and the dataset it was trained on in aa394e7.

“state-of-the-art visualizations”, by what metric?

Since we know of no quantiative metric that could settle this yet, and this is a subjective judgement, we changed the wording to "high-quality".

Would be good to state which objective(s) the paper is focusing on.

We have added that we are mostly using the channel objective.

Interesting that one neuron has these clusters of images at min/max activation (ie flowers and clocks). Could make it clear what “facets” means-- the kinds of images that result in different strength activations of the specific neuron? Or is it the kinds of images that result in maximum activation?

We mean the latter when we say facets and have added additional clarification for the term in the Diversity section.

Graphics: make it more clear that rightmost images = w/o diversity, and center images are with diversity

We have explicitly labelled those images as "Simple Optimization" and "Optimization with diversity" to make the difference clearer.

What diversity term is used in the examples?

We have added a footnote with the mathematical definition of our diversity term: cosine dissimilarity between the flattened Gram matrices.

A bit unclear on why optimization worked in above paragraphs, but didn’t over here?

We have added a section to the introduction of the optimization section stating that naive optimization doesn't work; linking to the section about challenegs in feature visualization by optimization.

Possibly include a section on future work in feature visualization? What high-level problems need to be solved?

We have added a section explicitly enumerating areas of future work we believe to be important.

Thank you again for your time and helpful comments! We think the article was significantly improved by incorporating your feedback. :-)