ReScience C · Machine Learning
[Re] A Novel CNN Ablation Reveals Limited Fourier Inductive Bias in LiteFNO:
A Reproducibility Study
We study LiteFNO, a lightweight Fourier Neural Operator for time-dependent PDEs,
in two ways: as a reproducibility study, since no public implementation of the
described architecture was available, and as a novel ablation that adds a
parameter-matched low-rank CNN the original paper never includes, isolating the
Fourier inductive bias. Our central question is whether the Fourier machinery
earns its complexity. The responsible-AI stake is reproducibility itself: an
architectural claim is worth only as much as the independent reimplementation
that can check it.
ReScience C · Medical imaging
[¬Re] Association of genomic subtypes of lower-grade gliomas with shape
features automatically extracted by a deep learning algorithm
We study the radiogenomic pipeline of Buda, Saha and Mazurowski (2019) in two ways:
as a reproducibility study, since no public implementation was available, and as a
stress test the original paper never runs, measuring how the reported association
between tumor shape and genomic subtype survives when it is computed on
automatically produced masks rather than manual ones. Our central question is
whether the shape features are stable enough to carry the claim. In a clinical
setting the answer decides whether a model’s output can responsibly be read as
evidence about a patient.
Sports analytics · Decision modeling
Challenge Run Value (cRV): quantifying strategic value in MLB’s ABS challenge
era
We study the Automated Ball-Strike challenge in two ways: as a run-expectancy
problem, pricing the state change a challenge produces against an empirically
derived surface built from public Statcast data, and as a resource-allocation
problem, since a challenge is scarce and spending one forecloses later options.
Our central question is what a challenge is worth at the moment it is spent. The
method is the argument: terminal outcomes that end a plate appearance have to be
priced rather than zeroed, and zeroing them inverts the sign of the headline
result, which is the kind of buried modeling choice a public, reproducible
pipeline exists to expose.
ReScience C · Medical imaging
[¬Re] Diffusion-Based Hierarchical Multi-Label Object Detection to Analyze
Panoramic Dental X-rays
We study HierarchicalDet, a diffusion-based detector trained across three tiers
of DENTEX panoramic X-ray annotation, in two ways: as a replication retrained
end-to-end on free-tier Kaggle hardware, since no trained weights are released,
and as an audit of what the public data can actually support. Our central
question is whether the pipeline can be checked by someone without a GPU
cluster. At our training budget the reproduction fails, reaching 6 to 13 percent
of the reported AP, but two findings survive any budget: the public test release
carries no Deep Caries label at all, so the reported diagnosis numbers cannot be
recomputed from public data by anyone, and the dataset ships under two different
licences at once.
ReScience C · Computational Biology
[Re] Improved survival analysis by learning shared genomic information from
pan-cancer data
We study VAECox (Kim et al., 2020), a transfer-learning survival model that
pretrains a variational autoencoder on pan-cancer RNA-seq before fine-tuning per
cancer, in two ways: as a reproduction retrained end-to-end on open-access TCGA
data, since the original cohort sits behind controlled access, and as an
extension the paper never runs, covering robustness, feature budgets,
lightweight models, subgroup fairness and risk stratification. Our central
question is whether transfer from pan-cancer data actually earns its keep. The
margins are the finding: VAECox keeps the best mean C-index but wins 5 of 10
cancers rather than the reported 7, and the per-cancer gaps are an order of
magnitude smaller than the seed-to-seed noise.
ReScience C · Machine Learning
[Re] MedMNIST v2: Replicating the ResNet baselines and auditing what macro-AUC
hides on DermaMNIST
We study the MedMNIST v2 ResNet baselines, which anchor a large body of follow-up
work, in two ways: as a replication from the paper’s description alone, without the
authors’ code and against a pre-registered tolerance, and as an equity audit of
DermaMNIST that the original never runs. Our central question is what the headline
metric hides. It is a fairness result with a clinical edge: macro-AUC is
essentially uncorrelated with class frequency while recall tracks it closely, so a
rare skin lesion can read as well classified and still go almost entirely
undetected.
Sports analytics · Player evaluation
Pitch-Quality Surplus (PQS): measuring hitter value net of pitch quality
We study hitter value in two ways: as a residual, subtracting a model of what each
pitch implied before contact from the run value the hitter actually produced, and as
a decomposition, splitting that surplus into what is earned against the hardest
pitches a hitter saw and what is earned against mistakes. Our central question is
what a batting line measures once the pitching is held constant. The two components
turn out to be nearly orthogonal, so how a hitter is good is two separable skills
that standard statistics report as one number.
AI × Humanities · Poetics
Restored Lines: A Model Finds the Poet’s Hardest Line Breaks and Declines to
Write Them
We study the line break, the one poetic decision no rule fixes, in two ways: by
stripping the lineation from 2,059 public-domain poems and asking a model to put
it back, graded by how much syntax each break severs, and by asking the same model
to write free verse of its own on matched subjects. Our central question is
whether recognising a creative decision and making one are the same capacity. The
gap between them says something practical about where these systems belong: closer
to an instrument for reading than a collaborator in writing.
ReScience C · Computational Biology
[Re] Robust deep learning–based protein sequence design using ProteinMPNN
We study ProteinMPNN in two ways: as a reproduction of its core sequence-design
results from the official code and weights, and as a robustness study the original
paper never runs, perturbing backbones with coordinate noise, masking residues, and
feeding deliberately malformed input. Our central question is what the model does
when the structure it is handed is wrong. The failure modes are the responsible-AI
finding: the more accurate model is not the more robust one, and most malformed
inputs fail silently rather than raising an error.
ReScience C · Machine Learning
[Re] TB-Net: A Tailored, Self-Attention Deep Convolutional Neural Network Design
for Detection of Tuberculosis Cases from Chest X-Ray Images
We study TB-Net (Wong et al., 2022), reported at 99.86% accuracy for tuberculosis
detection, in two ways: as a replication, which published materials do not permit,
since the architecture loads from a checkpoint whose download link is dead and the
training cohort has changed since publication, and as a compression and robustness
study aimed at smartphone deployment. Our central question is what survives the
move from a headline number to a clinic. It is a health-access result: the models
that would reach low-resource settings are the small ones, and we measure what
shrinking them costs.
AI × Humanities · Phonology
The Dead Ear: A Language Model’s Inherited Accent Is an Archive, Not an Author
We study inherited pronunciation in two ways: by building an item set from six dated
sound changes, so that “these words once rhymed” follows from a documented event
rather than a famous couplet, and by separating what a model reads off the page
from what it hears. Our central question is whether accents nobody speaks still
write. The answer bears on how we assign credit: a model can consult five
centuries of verse accurately and still keep it out of its own writing.
