submit

Submitting a paper to the workshop.

All submission must include a 1-page Storyline, as described here: https://jvgemert.github.io/storyline.pdf (examples included). Please follow the 1-page Storyline template available as a LaTeX or PDF file.

We have 2 tracks:

  • Storyline-only track: only the 1-page Storyline. This track is non-archival, so not part of any published proceedings, but authors can present a poster at the workshop for feedback and brainstorming.

  • Full paper track: 14 pages maximum (excluding references) in ECCV 2026 format with it’s 1-page Storyline in the appendix. Accepted papers will be part of the official workshop proceedings. Authors present a poster and we will select some work for oral presentation.

Timeline (any time on earth in 2026):

  • August 1: Final submission deadline (for both tracks).
  • August 7: Author acceptance notification.
  • August 10: Workshop author registration deadline.
  • August 14: Camera-ready deadline.

Submissions without a Storyline or a Storyline that does not follow the structure in https://jvgemert.github.io/storyline.pdf are at risk of desk rejection.

Not all elements of the Storyline always apply. Use the parts of the Storyline that are relevant to the research.

The Storyline will be used by reviewers to assess the work; see the reviewer questions below.

All submissions are anonymous, and full-papers must follow the ECCV 2026 Submission Policies.

Please note, we use Openreview, so if you don’t have an Openreview profile, please create one as soon as possible, and use an institutional email if possible; otherwise, approval might take up to 2 weeks.

Note: We might require reciprocal reviewing in case we get an unexpected high number of submissions.

Call for papers

Our workshop solicits contributions describing research work that explains why deep learning methods work, debunks or reinterprets existing common views, posits empirical regularities/laws, new empirical theories, etc.

Partly inspired by the Workshop on Scientific Methods for Understanding Deep Learning we consider the following topics a good fit:

  • Propose, validate and/or falsify hypotheses about the inner workings of deep networks,
  • Empirical observations to inform or inspire theoretical models,
  • Minimal analytical models that explain observed phenomena,
  • Controlled experiments for compiling rigorous empirical evidence,
  • Reproduce prior empirical results in simplified or extended settings,
  • introduce new experimental tools and methodologies for studying representation learning.

This is not meant as an exhaustive limited list and we aim for inclusion. We will interpret the scope broadly, and will err on the side of inclusion. Bold numbers are neither sufficient, nor necessary.

That said, we do aim for understanding, for theory, for empiricism, and for representation learning. Which means we do not aim for pure mathematical theory, and not about idiosyncratic systems nor data analysis.

Thus, the following type of papers are less fitting our scope:

  • Improving the state of the art without evidence why (we aim for understanding-based research);
  • Improving the state of the art without evidence that the improvements generalize (we aim for “theory”; ie: generalization beyond a single idiosyncratic system);
  • Pure mathematical machine learning theory paper (we aim for empirical theory);
  • “How well can Large Pre-trained AI Model X do task Y?” (We care less about finished artifacts with idiosyncratic training data, but instead aim for representation learning, so, with some “learning” involved).

Storyline: empirical rigor

The workshop will have a call for papers on empirical insights, empirical theory, how to do empirical representation learning research, meta-science for representation learning, etc. We are inspired by Hitchens’s razor: What can be asserted without evidence can also be dismissed without evidence and the Troubling Trends in ML scholarship paper (Lipton & Steinhardt, 2019) and others (Greydanus & Kobak, 2024); (Hung et al., 2025); (Picard, 2021); (Sculley et al., 2018) and design the review process accordingly, including questions such as

  • What is interesting about the paper?

  • Explanation vs Speculation: are all claims supported with empirical evidence?

  • Sources of empirical gains: what empirical evidence is there to support that the improved accuracy comes from what is claimed, and not due to some other confounding effect? (e.g. hyper-parameters?)

  • Methodological generalizability: how are the findings relevant to other methods/papers?

To accommodate empirical rigor, we will use a 1-page Storyline, to make empirical evidence explicit. The Storyline is described here: https://jvgemert.github.io/storyline.pdf (examples included). All papers need to offer a Storyline, we will also allow a Storyline-only submission, where poster boards are availble for early brainstorming. Additionally, we offer poster boards to relevant work from the main conference, offering authors an additional presentation venue.

Reviewer Questions

(Tentative) These are the questions we ask to reviewers on OpenReview.

1. Summary of the paper

Please give a short summary of the paper.

[Textbox]

2. Argumentation

See the 1-page Storyline (points 1-4) in the appendix of the submitted paper. Note, that not all parts of the Storyline need to be present, only the relevant parts. Consider the following questions.

  • What argumentation is there that the setting is reasonable/interesting? (Storyline: 1 Why Interesting?)
  • Are the related methods factual, relevant, complete? (Storyline: 2 How done now?)
  • Is the Problem and it’s consequences clearly described? (“So What?”) (Storyline: 3: Problem)
  • Is the Proposed improvement matching the problem? (Storyline: 4: Proposed)

[Textbox]

3. Empirical rigor/evidence (experiments)

See the 1-page Storyline: 5 Experiments in the appendix of the submitted paper. Note that not all experimental questions discussed in the Storyline need to be present, only the relevant experimental questions. Consider how is the following shown, and is it shown correctly:

  • Argumentation for the chosen datasets? (Storyline: c1, u1)
  • Baseline/competition methods are well reproduced/reasonable? (Storyline: c2, u2)
  • Baseline/competition methods suffering from the identified problem, and not some other confounder? (Storyline: c3, u3)
  • The proposed improvement come from what is claimed, and not due to some other confounding effect? (e.g. hyper-parameters?) (Storyline: c4, u4)

[Textbox]

4. Fit to the scope of the workshop

Is the paper a good fit to the goals of the workshop?

  • If the state of the art is improved, is there sufficient evidence for why?
  • How much empirical theory is there? Ie: how well would the paper’s insight generalize to other works?
  • Is it mainly mathematical theory without much empirical evidence?
  • Is it mainly an analysis of an existing (pre-trained) artifact without interventions?

[Textbox]

5. Strengths and weaknesses

What does the paper do well, and what can be improved.

[Textbox]

6. Conclusion

Reasons for the decision; feedback, suggestions for improvements, etc.

[Textbox]

7. Decision
  1. Clear accept
  2. Weak Accept
  3. Weak Reject
  4. Clear reject

 

 

Bibliography

2025

  1. Metascience for machine learning
    Hayley Hung, Marco Loog, and Jan Gemert
    2025
    Website

2024

  1. Scaling Down Deep Learning with MNIST-1D
    Samuel James Greydanus and Dmitry Kobak
    In International Conference on Machine Learning, 2024

2021

  1. Torch.manual_seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision
    David Picard
    ArXiv, 2021

2019

  1. Research for practice: troubling trends in machine-learning scholarship
    Zachary C Lipton and Jacob Steinhardt
    Communications of the ACM, 2019

2018

  1. Winner’s Curse? On Pace, Progress, and Empirical Rigor
    D. Sculley, Jasper Snoek, Alex Wiltschko, and Ali Rahimi
    2018