In this thesis, we combined the fields of computational image generation and incremental learning, with a focus on concept drift detection. We developed an innovative method for detecting conceptual changes in image data streams based on discriminator loss analysis within a generative adversarial network (GAN) learning loop. We conducted research that showed that the Wasserstein GAN architecture with gradient penalty (WGAN-GP), which we used in the main part of our testing, performed best for our purposes. We observed that when the distribution of the input data changes abruptly, the loss of the discriminator for a given number of iterations increases sharply in absolute value. We exploited this specific property in the development of the Gan Loss Drift Detection (GLDD) method. We thoroughly tested the method on an image dataset MNIST digits, which we had previously transformed into ten different distributions. During the experiments, the GLDD-KSWIN version performed particularly well, achieving an average precision of 0.79, a recall of 0.90, an F1-score of 0.84, and a latency of 41.64. The results show that the proposed method provides a promising foundation for further research in this area, which still remains largely untouched and challenging.
|
