Autonomous Control Uncertainty Estimation for Novel Image Inputs

The patent introduces an uncertainty module as an intermediary between the deep neural network decision module and the autonomous platform. This uncertainty module uses Sliced Wasserstein clustering to model the training data distribution and provides uncertainty estimates for novel data points. The clustering model acts as a mediator that bridges the gap between the decision-making capability of deep neural networks and the reliability needed for novel data, by identifying whether new data points fall within the learned data distribution before making decisions.

The patent changes the approach to parameter estimation by using Sliced Wasserstein distance instead of traditional maximum likelihood estimation with EM algorithm. This parameter change transforms the optimization problem into a different mathematical framework that avoids the local maxima problem inherent in EM algorithm. The Sliced Wasserstein clustering provides more accurate data distribution modeling by using optimal transport theory, which guarantees better convergence properties while maintaining computational feasibility through the slicing technique.

The patent extracts the clustering capability from the GAN framework and implements it as a separate Sliced Wasserstein clustering module. Instead of relying on the adversarial training process of GANs which is poor at clustering, the invention separates the data distribution modeling function from the synthetic image generation function. The Sliced Wasserstein clustering specifically targets and models the clustering structure of the training data, providing reliable uncertainty estimates, while synthetic images can be generated separately if needed without compromising the clustering accuracy.