Research map / Image generation / Autoregressive and latent generation
Learnware Specification via Dual Alignment
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 DALI, a learnware specification method that jointly applies discriminative alignment (H-discrepancy on pseudo-labels) and distribution alignment (MMD in random neural embedding spaces) to generate reduced-set specifications.
Problem
Existing learnware specification methods rely only on distribution alignment with the training data and ignore the model's discriminative performance, yielding inadequate characterization of the model's specialties.
Approach
Generate a class-labeled reduced specification by jointly minimizing an H-discrepancy term that aligns the model's pseudo-label performance with the specification and an MMD term in random neural embedding spaces that aligns class feature distributions with the true labels.
Key results
- In the homogeneous label space setting, DALI reaches Pre@1 of 100% with specification size K×5, matching RKME-W and above LANE (90.91%) and RKME (40.91-50%).
- In the mixed task setting with size K×10, DALI achieves quality 94.05% versus LANE's 90.23%, with superclass accuracy 100%.
- Membership inference attack accuracy on DALI specifications is 64.57%, comparable to RKME (66.13%) and LANE (64.00%), versus 90.55% for random samples.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add discriminative alignment (Ldis) to distribution alignment (ablation) | DALI · Homogeneous label space | Pre@1: 59.09 -> 100.00 | Helped |
| Add discriminative alignment (Ldis) to distribution alignment (ablation) | DALI · Homogeneous label space | Pre@2: 90.91 -> 100.00 | Helped |
| Add discriminative alignment (Ldis) to distribution alignment (ablation) | DALI · Heterogeneous label space | Pre@1: 59.09 -> 68.18 | Helped |
| Add discriminative alignment (Ldis) to distribution alignment (ablation) | DALI · Heterogeneous label space | Pre@2: 63.64 -> 72.73 | Helped |
| Add distribution alignment (LMMD) to discriminative alignment (ablation) | DALI · Homogeneous label space | Pre@1: 90.91 -> 100.00 | Helped |
| Add distribution alignment (LMMD) to discriminative alignment (ablation) | DALI · Homogeneous label space | Pre@2: 100.00 -> 100.00 | No clear effect |
| Add distribution alignment (LMMD) to discriminative alignment (ablation) | DALI · Heterogeneous label space | Pre@1: 63.64 -> 68.18 | Helped |
| Add distribution alignment (LMMD) to discriminative alignment (ablation) | DALI · Heterogeneous label space | Pre@2: 68.18 -> 72.73 | Helped |
| Use DALI instead of LANE | Learnware identification in heterogeneous label spaces · Heterogeneous label space | Pre@1: 63.64 -> 68.18 | Helped |
| Use DALI instead of LANE | Learnware identification in heterogeneous label spaces · Heterogeneous label space | Pre@2: 68.18 -> 72.73 | Helped |
| Use DALI instead of LANE | Learnware identification in mixed tasks · Mixed task setting | Superclass accuracy: 95.94 -> 100.00 | Helped |
| Use DALI instead of LANE | Specification quality in mixed tasks · Mixed task setting | Quality: 53.19 -> 75.68 | Helped |
Surprising findings
- Under the heterogeneous label space setting, DALI at size K×10 (63.64%) does not outperform the K×5 setting (68.18%) on Pre@1, and RKME-W matches DALI at size 100.
Limitations
- Privacy claim is supported mainly by a membership inference attack and heuristic arguments rather than formal guarantees.
- Theoretical bounds rely on simplifying assumptions (two-layer networks, single class) and the analysis is only partially extended to multi-layer networks.
- Experiments are limited to two image datasets with reconstructed label spaces and do not test other modalities.
Experimental setup
Datasets: DomainNet NICO
Benchmarks: DomainNet NICO
Metrics: Pre@k superclass accuracy quality (class accuracy of specifications) membership inference attack accuracy/ROC/precision
Baselines: RKME RKME-W LANE
Builds on: RKME specification LANE H-discrepancy Maximum Mean Discrepancy
Closest papers by meaning
- Handling Learnwares from Heterogeneous Feature Spaces with Explicit Label Exploitation (NeurIPS 2024 · similarity 0.85)
- Learnware Specification via Label-Aware Neural Embedding (AAAI 2025 · similarity 0.84)
- On the Ability of Developers' Training Data Preservation of Learnware (NeurIPS 2024 · similarity 0.84)
- Towards Making Learnware Specification and Market Evolvable (AAAI 2024 · similarity 0.83)
- Reasoning over Boundaries: Enhancing Specification Alignment via Test-time Deliberation (ICML 2026 · similarity 0.82)
- Slice-and-Pack: Tailoring Deep Models for Customized Requirements (AAAI 2025 · similarity 0.82)
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.