Cifar10 Histopathology Eval

cifar10-histopathology-eval

Jensen-Shannon Divergence Based Novel Loss Functions for Bayesian Neural Networks — Thiagarajan et al. (2022) (arXiv:2209.11366, 2022)

What this evaluates

Evaluates the generalization and uncertainty quantification of Bayesian Neural Networks trained with novel Jensen-Shannon divergence loss functions compared to standard KL divergence, specifically under noisy and class-biased data conditions.

Datasets

• CIFAR-10 — total ?; splits: train (-1), val (-1), test (-1)
• Breast Histopathology Dataset — total ?; splits: train (-1), val (-1), test (-1)

Metrics

• validation accuracy (primary) — range: [0, 1]
  • Proportion of correctly classified instances in the validation set, maximized during hyperparameter optimization.

Input / output format

Input: Normalized image data (CIFAR-10) or histopathology images, optionally augmented with varying levels of Gaussian noise.

Output: Class predictions or probability distributions from a ResNet-18 V1 Bayesian neural network.

Scoring recipe

def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return correct / len(gold_labels)

Common pitfalls

• Exact noise levels added to CIFAR-10 are not specified in the text.
• Hyperparameters are tuned via TPE/Hyperopt rather than fixed, making exact reproduction difficult without the appendix.
• Batch normalization layers are explicitly removed from the ResNet-18 architecture.

Evidence (verbatim from paper)

An optimization is performed to maximize the validation accuracy for different hyperparameter settings of the network.

Citation

@misc{thiagarajan2022jensen,
  title={Jensen-Shannon Divergence Based Novel Loss Functions for Bayesian Neural Networks},
  author={Thiagarajan et al. (2022)},
  year={2022},
  note={arXiv:2209.11366}
}

• arXiv: 2209.11366