Summary of estimation methods with associated cost function, regularizer and optimization technique.

Ridge, GLM and CbRF methods put a Gaussian prior on the filter coefficients, implemented by the -norm penalty in the objective function. The squared hinge loss of the CbRF method corresponds to the least squares loss with truncated negative part. Cross-validation on training data is used for all methods to determine the optimal regularization hyperparameter or the termination of early stopping for MID. STA: spike triggered average [19]. Ridge: ridge regression [67]. NRC: normalized reverse correlation [11]. CbRF: proposed classification-based receptive field estimation. GLM: generalized linear model [25], [26]. MID: maximally informative dimensions [18]. MI: mutual information between stimulus and response. SVD: singular value decomposition. CG: conjugate gradient. AUC: area under receiver operating characterisitc curve. SA: simulated annealing.