AI Paper Atlas

Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

Parameterless Transductive Feature Re-representation for Few-Shot Learning

Wentao Cui, Yuhong Guo

ICML 2021 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Alignment-based few-shot metric learning · 7 citations (OpenAlex)

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
apply parameterless re-representation (RR) at inference on top of inductive/meta-learning baselinesProtoNet, Matching, Relation (ResNet-18) · mini-ImageNetaccuracy: +2% to +4% (1-shot); +1.3% to +1.9% (5-shot)Improved
add SSL loss during meta-training (MT vs RR)ProtoNet, Matching, Relation · mini-ImageNetaccuracy: slight but consistent gains over RR versionImproved
apply RR at inference on SimpleShot, LaplacianShot, TIMSimpleShot, LaplacianShot, TIM · mini-ImageNet, tiered-ImageNet, CUB (1-shot and 5-shot)accuracy: SimpleShot+RR gains ~7-9% in 1-shot; TIM+RR >2% in all 1-shot tests; 5-shot gains marginalImproved
apply RR at inference on SimpleShot, LaplacianShot, TIMSimpleShot, LaplacianShot, TIM · mini-ImageNet, tiered-ImageNet, CUB 5-shotaccuracy: marginal gains (e.g., TIM+RR 85.04 to 85.20 on mini-ImageNet ResNet-18)No clear effect
remove query-to-query propagation (alpha1=0) (ablation)ProtoNet+MT, Matching+MT, Relation+MT, SimpleShot · mini-ImageNet 5-way 1-shotaccuracy: ProtoNet+MT 59.06 -> 57.86 when removedImproved
remove query-to-support propagation (alpha2=0) (ablation)ProtoNet+MT, Matching+MT, Relation+MT, SimpleShot · mini-ImageNet 5-way 1-shotaccuracy: ProtoNet+MT 59.06 -> 55.09 when removed; SimpleShot+RR 70.25 -> 64.14Improved
remove SSL loss during meta-training (ablation)ProtoNet+MT · mini-ImageNet 5-way 1-shotaccuracy: 59.06 -> 58.85 when removed (marginal)Improved
remove query-to-support propagation (alpha2=0) (ablation)LaplacianShot+RR and TIM+RR · mini-ImageNet 5-way 1-shotaccuracy: LaplacianShot+RR 75.04 -> 72.85 (alpha2=0 in 1-shot); TIM+RR 76.54 -> 76.00Mixed
remove query-to-query propagation (alpha1=0) (ablation)TIM+RR · mini-ImageNet 5-way 1-shotaccuracy: TIM+RR 76.54 -> 74.93Improved
apply multiple RR iterations at inference (beyond one) (ablation)TIM and SimpleShot · mini-ImageNet, tiered-ImageNet, CUB 1-shot and 5-shotaccuracyHurt
increase number of query shots (more unlabeled queries) with RRTIM+RR, SimpleShot+RR · 5-way 1-shot, mini-ImageNet and tiered-ImageNetaccuracyImproved
apply RR with increasing support shotsTIM+RR, SimpleShot+RR · 5-way, varying support shotsaccuracyMixed

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