Training Throughput
DriftXpress preserved sample quality close to standard drifting while substantially increasing throughput. Profiling on CIFAR10 found large reductions in field-estimation time for sharded and unsharded DriftXpress. Wall-clock convergence i…
1 sources - 4 claims
DriftXpress preserved sample quality close to standard drifting while substantially increasing throughput. Profiling on CIFAR10 found large reductions in field-estimation time for sharded and unsharded DriftXpress. Wall-clock convergence improved on SVHN, CIFAR10, and CIFAR100. Projected attraction is described as smoother and more stable because it uses a cached summary of the full positive support rather than mini-batch positives.