Currently EigenReporter supports streaming updates to its covariance statistics via the update method, but it does an unweighted average. For the training use-case, we'll want to use an exponential moving average since the representations of the model will change over time.