# LACS

Which examples receive backward computation during training no longer has to follow a preset quota or an absolute confidence cutoff. LACS compares each candidate's exponentiated maximum predictive confidence with the current population mean of exponentials, clips the coefficient at zero, and thereby applies a cutoff equal to a log-mean-exp reference plus a margin, with the retained count following the score distribution. On CIFAR-10 it removes 29% of backward rows relative to full cross-entropy at accuracy means of 89.77% versus 89.82%, with a three-seed interval crossing zero; on 14 tabular datasets quadratic boosting lowers mean test error from 23.36% to 22.96%.

Which examples receive backward computation during training no longer has to follow a preset quota or an absolute confidence cutoff. LACS compares each candidate's exponentiated maximum predictive confidence with the current population mean of exponentials, clips the coefficient at zero, and thereby applies a cutoff equal to a log-mean-exp reference plus a margin, with the retained count following the score distribution. On CIFAR-10 it removes 29% of backward rows relative to full cross-entropy at accuracy means of 89.77% versus 89.82%, with a three-seed interval crossing zero; on 14 tabular datasets quadratic boosting lowers mean test error from 23.36% to 22.96%.