AI Paper Atlas

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

Granularity-Aware Adaptive Classifier Expansion via Zero-Shot Learning

Xiangyu Wang, Yanze Gao, Changxin Rong, Lyuzhou Chen, Xiren Zhou, Huanhuan Chen

ICML 2026 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Generative feature synthesis for zero-shot

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
replace shared semantic-to-weight mapping with granularity-aware adaptive framework (CSD + ACS) (ablation)zero-shot classifier expansion · CUB (ZSL)top-1 accuracy T1 (%): 59.2 -> 68.2 (single semantics)Improved
add multi-source semantics generation (MSG) to a shared mapping (ablation)shared mapping baseline · CUB (ZSL)top-1 accuracy T1 (%): 59.2 -> 63.1Improved
add granularity modeling to multi-source semantics (ablation)MSG · AWA2 (ZSL)top-1 accuracy T1 (%): 62.6 -> 77.2Improved
add granularity-aware modeling vs richer semantics: average gains (ablation)2x2 attribution · CUB, AWA2, SUNaverage gain: semantics 3.8% vs granularity 8.3%Improved
add fine-grained adapters without structural priors (no CSD) (ablation)framework · ablation (Figure 3)T1 and HImproved
introduce CSD module as routing priors (ablation)framework · ablation (Figure 3)T1 and HImproved
add causal invariance loss and multi-source semantics in ACS (ablation)framework · ablation (Figure 3)T1 and HImproved
give baselines the MSG semantics as inputICIS · CUB (GZSL)harmonic mean H (%): 57.2±1.0 -> 59.4±0.9Improved
give baselines the MSG semantics as inputICIS · SUN (GZSL)harmonic mean H (%): 35.1±2.9 -> 33.6±2.5Hurt
give baselines the MSG semantics as inputCEMIL · CUB (GZSL)harmonic mean H (%): 60.7±0.7 -> 60.9±0.8No clear effect
use proposed method instead of CEMIL under distribution shift from ImageNetzero-shot classifier expansion · AWA2 (Wiki2Vec, ZSL)top-1 accuracy T1 (%): 73.3±1.2 -> 83.1±0.9Improved
use proposed method in class-incremental expansionzero-shot classifier expansion · AWA2 (5 steps)harmonic mean H (%): 59.7±2.5 -> 66.7±2.1 at step 5Improved

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