The Proximity of the Inception Score as an Evaluation Criterion

Generative Adversarial Networks (GANs) require specific metrics to evaluate the quality and diversity of the synthetic images they produce. The Inception Score (IS) is a quantitative metric for this purpose, using a pre-trained Inception network to assess if generated images are both high-quality and varied. IS is calculated by measuring the KL divergence between the conditional probability of an image belonging to a class and the marginal probability across all generated images. A good model produces images with low-entropy conditional probabilities (high confidence) but a high-entropy marginal probability (high diversity). However, the metric has limitations, as its effectiveness is reduced when the generated images belong to classes not present in the ImageNet dataset on which the Inception model was trained.