AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Learning To Stop While Learning To Predict

Xinshi Chen, Hanjun Dai, Yu Li, Xin Gao, Le Song

ICML 2020 · Architectures and efficiency › RNNs and normalization · approach: Dataset distillation and data-efficient training · 18 citations (OpenAlex)

Read the paper

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
add an input-specific stopping policy to LISTALISTA · Sparse recoveryNMSE (dB): -17.53 -> -22.41 (mixed SNR)Improved
add task-specific stopping of adaptation stepsMAML · Task-imbalanced Omniglot, 20-way 1-5 shotaccuracy: 97.96 ± 0.3% -> 98.45 ± 0.2%Improved
add task-specific stopping of adaptation stepsMAML · Task-imbalanced MiniImagenet, 5-way 1-10 shotaccuracy: 57.20 ± 1.1% -> 60.67 ± 1.0%Improved
add task-specific stopping of adaptation stepsMAML · Vanilla few-shot meta learning, Omniglot and MiniImagenetaccuracy: Omniglot 5-way 1-shot: 98.7 ± 0.4% -> 99.62 ± 0.22%; MiniImagenet 5-way 5-shot: 63.11 ± 0.92% -> 63.41 ± 0.80%Mixed
add an input-specific stopping policy to DnCNNDnCNN · BSD68, noise σ=35, 45, and 55PSNR: 27.60 -> 27.61 (σ=35); 26.56 -> 26.59 (σ=45); 25.71 -> 25.79 (σ=55)Improved
add an input-specific stopping policy to DnCNNDnCNN · BSD68, unseen noise σ=65 and σ=75PSNR: 22.19 -> 23.56 (σ=65); 17.90 -> 18.62 (σ=75)Improved
use oracle stopping after Stage I trainingVGG16 · Tiny-ImageNetaccuracy: 58.60% -> 83.26%Improved
replace oracle stopping with the learned sequential policyVGG16 · Tiny-ImageNetaccuracyHurt
replace two-stage oracle training with joint AEVB optimization (ablation)LISTA-stop · Sparse recovery, mixed SNRNMSE (dB): -22.41 -> -21.92Hurt
add optional joint fine-tuning after Stage I + II (ablation)LISTA-stop · Sparse recovery, mixed SNRNMSE (dB): -22.41 -> -22.78Improved
use forward KL rather than reverse KL for policy imitation (ablation)Stopping policy · Experiments across tasksprediction performanceImproved

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