AI Paper Atlas

Research map / Optimization / Optimizers

Active Bias: Training More Accurate Neural Networks by Emphasizing High Variance Samples

Haw-Shiuan Chang, Erik Learned-Miller, Andrew McCallum

NeurIPS 2017 · Optimization › Optimizers · approach: SGD momentum and step-size schedules

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Emphasize samples by prediction variance rather than uniform samplingSGD training · CIFAR 10testing error rate (%): 62.49$\pm$0.06 -> 60.66$\pm$0.06Improved
Emphasize samples by prediction variance rather than uniform samplingSGD training · Noisy MNISTtesting error rate (%): 0.83$\pm$0.01 -> 0.64$\pm$0.01Improved
Emphasize samples by threshold closeness rather than uniform samplingSGD training · Noisy MNISTtesting error rate (%): 0.83$\pm$0.01 -> 0.63$\pm$0.01Improved
Emphasize samples by prediction variance rather than uniform samplingSGD training · MNISTtesting error rate (%): 0.55$\pm$0.01 -> 0.51$\pm$0.01Improved
Weight samples by prediction variance instead of uniform scanSGD-Scan · CoNLL 20031-F1 error rate (%): 11.62$\pm$0.04 -> 11.24$\pm$0.06Improved
Weight samples by threshold closeness instead of uniform scanSGD-Scan · OntoNote 5.01-F1 error rate (%): 17.80$\pm$0.05 -> 17.51$\pm$0.05Improved
Weight samples by easiness rather than uniform scanSGD-Scan · CIFAR 100, ResNet-63testing error rate (%): 30.70$\pm$0.06 -> 29.92$\pm$0.09Improved
Weight samples by difficulty rather than uniform scanSGD-Scan · Noisy MNISTtesting error rate (%): 0.81$\pm$0.01 -> 0.92$\pm$0.01Hurt
Sample by difficulty rather than uniform samplingSGD training · Noisy MNISTtesting error rate (%): 0.83$\pm$0.01 -> 1.00$\pm$0.01Hurt
Use threshold-closeness weighting instead of prediction-variance weighting (ablation)SGD-Scan · CIFAR 10testing error rate (%): 60.61$\pm$0.06 vs 61.02$\pm$0.06No clear effect
Apply threshold-closeness weighting in addition to distillationDistillation training · MNIST (distill)testing error rate (%): 2.27$\pm$0.01 -> 2.07$\pm$0.02Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy