Derivative-free optimization is required, whenever gradients of objective functions are not available or too expensive to evaluate. A typical imaging application, where such a situation arises, is when employing a closed-box classification tool, like a neural network, where only the model output is available. Consensus-based techniques utilize an ensemble of particles exploring a state space, while simultaneously using function evaluations to update their guess of a global minimizer. This strategy is not restricted to optimization but can also be used for sampling.

CBX provides Python and Julia packages that offer a general, high-level interface for consensus-based optimization and sampling.