Research map / Label-efficient and robust learning / Recognition and long-tail

Learnware Specification via Label-Aware Neural Embedding

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 1 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 LANE, a label-aware learnware specification built by matching class-wise feature distributions in a random neural embedding space instead of an RKHS, and evaluates it on CIFAR100 and 20newsgroups model reuse.

Problem

Learnware specification generation via RKME relies on feature-only information and on kernel choice, which limits how well specifications characterize model specialty for identifying reusable learnwares.

Approach

Replace the RKHS with a neural embedding space built from numerous random neural networks, and use labels to match intra-class feature distributions via MMD, so the specification is kernel-free and label-aware; deployment uses class-wise mixture weight estimation via quadratic programming and mimic data sampling.

Key results

• On CIFAR100 with 1 mixed task, LANE reaches 84.50% accuracy vs. 84.04% for RKME and 51.69% for MAX.
• On 20newsgroups with 1 mixed task, LANE achieves 74.29% vs. 73.83% for RKME, but RKME is slightly better at 3-5 mixed tasks in part of the comparison (e.g., 64.78 vs 64.28 at 5 tasks).
• At 20 mixed tasks on CIFAR100, LANE achieves 68.13% vs. 66.27% for RKME.

Measured effects reported in the paper

Surprising findings

• LANE's advantage over RKME is small and reverses at larger numbers of mixed tasks on CIFAR100 (10 tasks: 72.51 vs 73.02) and 20newsgroups (5 tasks: 64.28 vs 64.78).

Limitations

• Performance depends on specification size and degrades as the number of mixed tasks grows relative to RKME in some settings.
• Gains over RKME are modest and not uniform across mixing settings.
• Evaluated only on two superclass-structured benchmarks with a fixed global feature extractor.

Experimental setup

Datasets: CIFAR100 20newsgroups

Benchmarks: CIFAR100 20newsgroups

Metrics: accuracy

Baselines: MAX HMR RKME

Builds on: Reduced Kernel Mean Embedding learnware paradigm Dataset distillation with random networks (Dong, Zhao, and Lyu 2022) HMR

Closest papers by meaning

• On the Ability of Developers' Training Data Preservation of Learnware (NeurIPS 2024 · similarity 0.89)
• Handling Learnwares from Heterogeneous Feature Spaces with Explicit Label Exploitation (NeurIPS 2024 · similarity 0.89)
• Learnware Specification via Dual Alignment (ICML 2025 · similarity 0.84)
• UME-R1: Exploring Reasoning-Driven Generative Multimodal Embeddings (ICLR 2026 · similarity 0.84)
• Neural-Kernel Conditional Mean Embeddings (ICML 2024 · similarity 0.83)
• Breaking the Batch Barrier (B3) of Contrastive Learning via Smart Batch Mining (NeurIPS 2025 · similarity 0.83)