Drifting Models
Standard drifting avoids adversarial discriminator training but shifts substantial computation into kernel-based field estimation during training. Drifting models generate samples with a single generator evaluation. Their training signal u…
1 sources - 4 claims
Standard drifting avoids adversarial discriminator training but shifts substantial computation into kernel-based field estimation during training. Drifting models generate samples with a single generator evaluation. Their training signal uses attraction toward data and repulsion away from the current model distribution in feature space. The exact drifting field is anti-symmetric and becomes zero when the data and model distributions match.