There is growing evidence that human and animals represent the uncertainty associated with sensory stimuli and utilize this uncertainty during action planning and decision making in a statistically optimal way. Recently, a nonparametric framework of representing such probabilistic information has been proposed where neural activity encodes samples from the distribution at every moment. I will review the empirical and theoretical motivations and details of such a sample-based probabilistic representations in the cortex, including two major issues that need to be clarified to view the framework as a viable alternative for cortical computation. First, is there sufficient time to collect enough samples for an accurate estimate of a stimulus, especially considering that the natural environment is in constant change? Second, can biases in the approximation during learning due to a small number of samples be avoided? We explored these issues by comparing a neural circuit performing cue combination using a sampling-based representation to an optimal estimator. I will report results suggesting that sample-based representations are highly feasible alternatives for performing complex hierarchical probabilistic cortical computations in high-dimensional spaces.