Figure 6
A comparison of gradients computed using different approximation schemes for t-based noise with ν = 3. (a)–(c) and (d)–(f) depict the behavior of the gradient approximations with a decreasing noise level. Gradients were computed using a maximum-likelihood estimate of σA using their corresponding approximations. Both the normal model and the t-based model clearly outperform the French–Wilson and the Sivia approaches, while a marginal improvement over the normal noise model is obseved when the t-based model is used. |