# ReScience C submission metadata.
# Filled where the repository already determines the answer.

# Title of the article:
#  - "[Re]" prefix = successful replication, "[¬Re]" = failed replication.
#  - We use [Re]: the central claims of the target paper reproduce. Two absolute
#    numbers undershoot and one supporting Chen claim does not reproduce; those are
#    reported as partial results in the paper, not as a failed replication. If the
#    editors read the MNIST NODE undershoot as a failure, this prefix must change.
title: "[Re] Augmented Neural ODEs"

# List of authors with name, orcid number, email and affiliation.
# Affiliation "*" means contact author (required even for single-authored papers).
# Authors are in alphabetical order by surname and contributed equally (see content.tex).
authors:
  - name: Giuseppe Bassi
    orcid: 0009-0009-5441-6128
    email: giuseppe.bassi@princeton.edu
    affiliations: 1

  - name: Kabir Ben Nefla
    orcid: 0009-0000-7466-3091
    email: benneflakabir@gmail.com
    affiliations: 1

  - name: Gaia Grossi
    orcid: 0009-0004-3192-1678
    email: arteniagrossi@gmail.com
    affiliations: 1

  - name: Anna Notaro
    orcid: 0009-0006-6412-2164
    email: annalytonotaro@gmail.com
    affiliations: 1

  - name: Beatrice Porta
    orcid: 0009-0005-7555-2621
    email: beatrice.porta04@gmail.com
    affiliations: 1

  - name: Leonardo Vanni
    orcid: 0009-0003-1869-9286
    email: leonardo.vanni@tum.de
    affiliations: 1,*            # * is for contact author

# List of affiliations with code (corresponding to author affiliations), name
# and address.
affiliations:
  - code:    1
    name:    Bocconi University
    address: Milan, Italy

# List of keywords (adding the programming language might be a good idea)
keywords: rescience c, python, pytorch, neural ordinary differential equations, augmented neural odes, torchdiffeq

# Code URL and DOI/SWH (url is mandatory for replication, doi after acceptance).
# The concept DOI is cited rather than the DOI of one release, so that a later release
# does not leave it stale. The exact snapshot submitted is recorded in the ReScience C
# submission issue, together with its version DOI.
code:
  - url: https://github.com/VanniLeonardo/re-augmented-neural-odes
  - doi: 10.5281/zenodo.22704199   # concept DOI: always resolves to the newest release
  - swh:

# Data URL and DOI (optional if no data)
# MNIST and CIFAR-10 are standard public datasets fetched by the data loaders at
# run time; no data artifact is deposited. Per-seed result CSVs live in the code
# repository under results/ and are covered by the code DOI above.
data:
  - url:
  - doi:

# Information about the original article that has been replicated
replication:
 - cite: "Emilien Dupont, Arnaud Doucet and Yee Whye Teh. Augmented Neural ODEs.
          Advances in Neural Information Processing Systems 32 (NeurIPS 2019)."
 - bib:  dupont2019augmented
 - url:  https://arxiv.org/pdf/1904.01681
 - doi:  10.48550/arXiv.1904.01681

# Don't forget to surround abstract with double quotes
abstract: "We replicate Augmented Neural ODEs (Dupont et al., 2019), reimplemented from
           the paper rather than from the authors' code, together with four supporting
           claims about solver cost and adjoint memory from Neural Ordinary Differential
           Equations (Chen et al., 2018). The structural claims reproduce: a Neural ODE
           cannot represent the one-dimensional crossing map and fails at exactly the error
           floor the accompanying proposition predicts; its integration cost grows during
           training where an augmented model's stays flat; and it generalises poorly into a
           region withheld from training, which augmentation repairs at every width and on
           both geometries we tested. At matched parameter counts the augmented model is
           more accurate and cheaper on MNIST and CIFAR-10, although two of the four
           absolute accuracies fall short of the published values. Of the four supporting
           claims, two reproduce, one reproduces partially, and one does not: the claim that
           the backward pass costs about half the forward pass in function evaluations. With
           code shared by the original authors we traced their value to a counting artefact.
           The corrected value of about 0.8 remains far from ours, a difference we could not
           attribute to the tolerance, the error norm or the solver. We attach an invertibility check to every measurement. An adaptive solver at
           too loose a tolerance returns a plausible trajectory and a plausible evaluation
           count while never tracking the flow, and the vector field of a Neural ODE stiffens
           as it trains. Several conclusions we reached during this work were retracted for that
           reason."

# Bibliography file (yours)
bibliography: bibliography.bib

# Type of the article
# Type can be: Editorial / Letter / Replication
type: Replication

# Scientific domain of the article (e.g. Computational Neuroscience)
domain: Machine Learning

# Coding language (main one only if several)
language: Python


# To be filled by the author(s) after acceptance
# -----------------------------------------------------------------------------

# For example, the URL of the GitHub issue where review actually occured
review:
  - url:

contributors:
  - name:
    orcid:
    role: editor
  - name:
    orcid:
    role: reviewer
  - name:
    orcid:
    role: reviewer

# This information will be provided by the editor
dates:
  - received:
  - accepted:
  - published:

# This information will be provided by the editor
article:
  - number:
  - doi:
  - url:

# This information will be provided by the editor
journal:
  - name:   "ReScience C"
  - issn:   2430-3658
  - volume:
  - issue: