Research map / Label-efficient and robust learning / Noisy labels

Robust Self-Paced Hashing for Cross-Modal Retrieval with Noisy Labels

AAAI 2025 · Label-efficient and robust learning › Noisy labels · approach: Noisy cross-modal hashing and multi-label · 17 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RSHNL, a cross-modal hashing method that uses self-paced learning with a per-pair difficulty weight to identify and drop noisy-label pairs while training from easy to hard, combined with contrastive hashing and hash-center aggregation losses.

Problem

Cross-modal hashing degrades when training labels are noisy; existing self-paced hashing methods address feature noise or outliers but not label noise.

Approach

Uses a self-paced regularizer (linear interpolation, weight w_i = max(0, 1 - loss/gamma)) so pairs whose loss exceeds the pace threshold are treated as noisy and given zero weight, while clean pairs are learned from easy to hard; combined with a contrastive hashing loss (CHL) for cross-modal consistency and a center aggregation loss (CAL) toward class hash centers.

Key results

• On XMedia I2T at 128 bits with 0.2 noise, RSHNL reaches 90.8 MAP versus 90.5 for the next best DHRL.
• On INRIA-Websearch I2T at 128 bits with 0.8 noise, RSHNL reaches 42.9 MAP versus 38.5 for DHaPH.
• Ablation on INRIA-Websearch at 0.8 noise: removing SPL drops 128-bit MAP to 20.2 while full RSHNL gets 42.9.

Measured effects reported in the paper

Surprising findings

• Removing the warm-up training (RSHNL-1) yields flat MAP across all noise rates (84.3 on XMedia), indicating an unexpected dependence on warm-up.
• Several baselines (e.g., DHRL, CMMQ) show sharp non-monotonic or collapsing performance as noise and bit length change, and some methods run out of memory on XMediaNet.

Limitations

• Performance degrades with more categories (INRIA-Websearch, XMediaNet), where most methods fail.
• Selecting the learning pace parameter gamma requires sensitivity analysis and is bounded by M(r^2-r+1)/r.
• Evaluation uses only symmetric synthetic label noise.

Experimental setup

Datasets: XMedia INRIA-Websearch Wikipedia XMediaNet

Benchmarks: XMedia INRIA-Websearch XMediaNet

Metrics: MAP precision-recall curves

Baselines: DGCPN CIRH UCCH WASH HCCH DSCMH HMAH CMMQ MIAN DHRL DHaPH

Builds on: self-paced learning DHaPH DSCMH CMMQ

Closest papers by meaning

• Dual Self-Paced Cross-Modal Hashing (AAAI 2024 · similarity 0.95)
• Robust Self-reflective Hashing for Cross-modal Retrieval with Noisy Label (ICML 2026 · similarity 0.93)
• Exploiting Descriptive Completeness Prior for Cross Modal Hashing with Incomplete Labels (NeurIPS 2024 · similarity 0.89)
• Distribution-Consistency-Guided Multi-modal Hashing (AAAI 2025 · similarity 0.89)
• Lightweight Contrastive Distilled Hashing for Online Cross-modal Retrieval (AAAI 2025 · similarity 0.88)
• Deep Evidential Hashing for Trustworthy Cross-Modal Retrieval (AAAI 2025 · similarity 0.88)