Team ProjectComputer Vision · Class Imbalance · Deep Learning

Long-Tail Deep Learning

A small number of classes contain many examples, while thousands of rarer classes have very little training data. The work tests techniques that stop overall accuracy from hiding failure on rare species.

Central findingLogit adjustment produced the strongest balanced result on the documented iNaturalist experiment.

Evidence

The result in context

CIFAR-10 accuracy: 96.20%
iNaturalist species: 8,142
iNaturalist images: 461,939
Best balanced accuracy: 58.72%

Question

Which training adjustments improve balanced performance when class frequencies differ by roughly 200 to one?

A team study of image classification when a few classes have abundant examples and thousands of rarer classes have very little data.

Head, medium and tail species accuracy across long-tail training methods — Logit adjustment produces the strongest balanced result, improving rare-species accuracy without hiding the head–tail trade-off.

Problem

Overall accuracy can hide rare-class failure

A small number of classes contain many examples, while thousands of rarer classes have very little training data. On iNaturalist, 8,142 species and 461,939 images produce approximately 200:1 class imbalance.

A model can score well by favouring common classes, so balanced accuracy gives each class more equal influence on the result.

iNaturalist class-frequency distribution on a logarithmic scale — A small number of species have many training images while thousands of tail species have far fewer examples.

Experiments

Training objectives were adjusted for the long tail

The team compared approaches designed to correct class-frequency bias. Logit adjustment produced the strongest balanced result: 58.72% balanced accuracy and 56.19% tail accuracy.

A separate CIFARNet implementation reached 96.20% CIFAR-10 accuracy, providing a conventional balanced-dataset reference.

Trade-off between common-class and rare-class accuracy — The objective is not simply higher overall accuracy, but a better balance between common and rare classes.

Limitations

What this evidence does not establish

Results are tied to the documented splits, architecture and training setup.
Balanced accuracy improves visibility of tail performance but does not resolve every deployment cost or label-quality issue.
The pinned README reports 58.72% balanced and 56.19% tail accuracy, while stored notebook output rounds to 58.73% and 56.21%. Public copy follows the locally retained full-precision CSV at 58.72% and 56.19%; that CSV and the plotted outputs are not present in the pinned public commit.

Source and reproducibility

Trace the evidence

Source code and the public notebook are available in the repository. The final full-precision analysis CSV and plotted outputs are retained as local project evidence and are not present in the pinned public commit.

View repository

Final public team repositoryBeyond_CIFAR_10_image_classification.ipynbCommit / evidence ID: ddacc7132968ab12c8586c81bec02e365861d455
Locally retained full-precision analysis output (not in pinned public commit)analysis/results_summary.csv