AI-ordlista
Den kompletta ordlistan över AI
Generative Adversarial Network
Unsupervised learning architecture composed of two competing neural networks, a generator and a discriminator, that compete against each other to generate realistic synthetic data.
Minimax Loss
Original objective function of GANs where the generator minimizes the log-probability of the discriminator being wrong, while the discriminator maximizes the probability of correct classification.
Latent Space
Reduced dimensional vector space where the generator samples random noise to create data, allowing semantic control over the generated features.
StyleGAN
Advanced GAN architecture using a mapping network and AdaIN modules to control hierarchical styles of generated features at different resolutions.
Jensen-Shannon Distance
Symmetric and bounded divergence metric used in original GANs to measure the difference between real and generated data distributions.
Gradient Penalty
Regularization term added to WGAN loss function to constrain discriminator gradients, ensuring the continuity of the Lipschitz transform.
Nash Equilibrium
Optimal point where neither the generator nor the discriminator can improve their performance by unilaterally modifying their parameters, indicating training convergence.
Encoder Network
Additional component in BiGAN or ALI variants that learns to map real data to latent space, enabling latent inference and reconstruction.
Cycle Consistency Loss
Additional loss function in CycleGANs ensuring content preservation during translations between unpaired domains via back-and-forth cycles.
Spectral Normalization
Regularization technique constraining the spectral norm of discriminator weights, stabilizing GAN training by controlling the Lipschitz transform.
Progressive Growing
Training strategy where networks start with low-resolution images and progressively add layers to increase resolution, stabilizing convergence.
Variational Auto-Encoder
Hybrid architecture combining VAE and GAN where the VAE ensures latent space coverage and the GAN improves visual quality of generated samples.
Fréchet Inception Distance
Quantitative evaluation metric measuring the similarity between Inception feature distributions of real and generated images via the Fréchet distance.