Inception-Score
Das Gründung Score (IS) is a metric used to evaluate the quality of images generated by künstliche Intelligenz models, particularly generative adversarial networks (GANs). It provides a numerical score that reflects both the clarity and diversity of the generated images.
Der IS wird aus einem vortrainierten Convolutional Neural Network, typically the Inception v3 model, which classifies images into various categories. The score is calculated in two main steps:
- Klarheit: For each generated image, the model predicts a probability distribution over the possible classes. A higher score indicates that the image is more confidently classified into a specific category, suggesting greater clarity.
- Vielfalt: The score also considers the diversity of generated images by evaluating how different the images are from one another. It does this by looking at the distribution of predicted classes across multiple generated images. A good model should produce images that belong to a wide range of categories, thereby achieving high diversity.
Mathematisch wird der Inception-Score wie folgt berechnet:
IS = exp(E[KL(p(y|x) || p(y))])
wobei:
- p(y|x) is the bedingte Wahrscheinlichkeit modelliert von Klassenlabels gegeben ein Bild x voraus.
- p(y) is the Marginalwahrscheinlichkeit der Klassenlabels über alle generierten Bilder hinweg.
- KL represents the Kullback-Leibler-Divergenz, which measures how one probability distribution diverges from a second, expected probability distribution.
Overall, a higher Inception Score indicates that the AI model is generating images that are both high in quality (clear and recognizable) and diverse, making it a valuable tool for researchers and developers in the field of Computer Vision und generatives Modellieren.