Cifar100 Miniimagenet Eval

cifar100-miniimagenet-eval

VDLF-Net: Variational Feature Fusion for Adaptive and Few-Shot Visual Learning — Yan (2026) (arXiv:2604.23641, 2026)

What this evaluates

This evaluation protocol probes a model's ability to perform standard supervised image classification and adaptive few-shot episodic learning. It measures how well the model generalizes to unseen classes under limited supervision by averaging accuracy over multiple sampled episodes.

Datasets

• CIFAR-100 — total 60000; splits: train/test (-1)
• Mini-ImageNet — total 60000; splits: train (38400), val (9600), test (12000)

Metrics

• accuracy (primary) — range: [0, 1]
  • For standard classification, accuracy is the fraction of correctly predicted images out of the total test set. For few-shot learning, it is the mean accuracy across 100 test episodes, where each episode's accuracy is the fraction of correctly classified query images. 95% confidence intervals are computed using a normal approximation: ±1.96 * σ / √100.
• macro precision/recall/F1 — range: [0, 1]
  • Standard macro-averaged precision, recall, and F1-score computed across all 100 classes for the CIFAR-100 supervised task.

Input / output format

Input: RGB images (32×32 for CIFAR-100; standard resolution for Mini-ImageNet). For few-shot episodes, inputs are provided as a support set (N×K images) and a query set (N×Q images).

Output: Predicted class label or probability distribution over the N classes for each input image.

Scoring recipe

# Standard classification
correct = sum(pred == gold for pred, gold in zip(predictions, golds))
accuracy = correct / len(golds)

# Few-shot episodic
episode_accuracies = []
for episode in test_episodes:
    q_correct = sum(pred == gold for pred, gold in zip(episode.query_preds, episode.query_golds))
    episode_accuracies.append(q_correct / len(episode.query_golds))
mean_acc = sum(episode_accuracies) / len(episode_accuracies)
ci = 1.96 * np.std(episode_accuracies) / np.sqrt(len(episode_accuracies))

Common pitfalls

• Using data augmentation (e.g., random crops/flips) during evaluation instead of the specified center crop.
• Reporting single-run few-shot accuracy without averaging over 100 test episodes and computing the 95% confidence interval.
• Confusing the standard supervised train/test split with the 64/16/20 episodic split used for Mini-ImageNet.

Evidence (verbatim from paper)

CIFAR-100: accuracy and macro precision/recall/F1. Few-shot: episode mean accuracy with normal-approximation 95% intervals (100 test episodes).

Citation

@misc{yan2026vdlfnet,
  title={VDLF-Net: Variational Feature Fusion for Adaptive and Few-Shot Visual Learning},
  author={Yan (2026)},
  year={2026},
  note={arXiv:2604.23641}
}

• arXiv: 2604.23641