After neural network training, one can find a more appropriate stationary point of the linearized model or the last layer as proposed in Sec. 3.2 here. This can improve the Laplace approximation and should generally be cheap for last layers. Implementation-wise, this could be an option for .fit(refine=True) since the data loader is required.
After neural network training, one can find a more appropriate stationary point of the linearized model or the last layer as proposed in Sec. 3.2 here. This can improve the Laplace approximation and should generally be cheap for last layers. Implementation-wise, this could be an option for
.fit(refine=True)
since the data loader is required.