Research map / Label-efficient and robust learning / Domain adaptation and generalization
Incorporating probabilistic domain knowledge into deep multiple instance learning
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 DeeMILIP, three strategies (direct prior injection, attention modulation, attention training) to inject noisy instance-level probabilistic domain knowledge into attention-based deep multiple instance learning, evaluated on semi-synthetic immune repertoire data and MNIST-MIL.
Problem
Deep learning models have limited ability to incorporate quantitative domain knowledge because their components rarely map to domain entities; in MIL with low witness rates and small datasets, learning good attention is hard.
Approach
MIL's structure lets attention outputs be tied to instance-level class probabilities, so noisy external instance priors can be injected directly into the attention output (DPI, AM) or used as an auxiliary instance-level cross-entropy loss (AT), without adding trainable parameters.
Key results
- At 0.05% witness rate, AM reaches ROCAUC 0.84 vs 0.53 for Vanilla, allowing learning at a twenty-fold lower witness rate than Vanilla (which needs 1%).
- With 12 training bags at 2% witness rate, AM achieves ROCAUC 0.98 versus 0.56 for Vanilla, which needs 360 bags to perform well (thirty-fold smaller dataset).
- On MNIST-MIL with attention pooling at 0.5% witness rate, AT reaches ROCAUC 0.92 vs 0.57 for Vanilla.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| modulate learned attention with instance priors | DeepRC · Immune diagnostics, WR 0.05% | ROCAUC: 0.53 ± 0.02 -> 0.84 ± 0.05 | Helped |
| modulate learned attention with instance priors | DeepRC · Immune diagnostics, WR 0.2% | ROCAUC: 0.54 ± 0.01 -> 0.99 ± 0.00 | Helped |
| modulate learned attention with instance priors | DeepRC · Immune diagnostics, WR 2%, training size 12 | ROCAUC: 0.56 ± 0.03 -> 0.98 ± 0.02 | Helped |
| modulate learned attention with instance priors | DeepRC · Immune diagnostics, WR 2%, 1000 updates | ROCAUC: 0.59 ± 0.06 -> 1.00 ± 0.00 | Helped |
| add instance priors as fixed attention weights | DeepRC · Immune diagnostics, WR 2% | ROCAUC: 0.75 ± 0.07 -> 0.93 ± 0.07 | Helped |
| add learned attention to prior injection | DeepRC · Immune diagnostics, WR 2% | ROCAUC: 0.93 ± 0.07 -> 1.00 ± 0.00 | Helped |
| add an auxiliary instance-level attention loss | DeepRC · Immune diagnostics, WR 2% | ROCAUC: 0.81 ± 0.12 -> 1.00 ± 0.00 | Helped |
| add attention learning without instance priors | DeepRC average pooling · Immune diagnostics, WR 2% | ROCAUC: 0.75 ± 0.07 -> 0.81 ± 0.12 | Helped |
| modulate learned attention with instance priors | Attention pooling · MNIST-MIL, WR 0.5% | ROCAUC: 0.57 ± 0.05 -> 0.55 ± 0.08 | No clear effect |
| add an auxiliary instance-level attention loss | Attention pooling · MNIST-MIL, WR 0.5% | ROCAUC: 0.57 ± 0.05 -> 0.92 ± 0.02 | Helped |
| use instance priors as fixed attention weights | Attention pooling · MNIST-MIL, WR 0.5% | ROCAUC: 0.57 ± 0.05 -> 0.00 ± 0.00 | Hurt |
| modulate learned attention with instance priors | Attention pooling · MNIST-MIL, WR 3% | ROCAUC: 0.87 ± 0.07 -> 1.00 ± 0.00 | Helped |
Surprising findings
- Attention modulation improves instance-level attention on low-prior witnesses even though it is not trained at instance level.
- Under the MNIST-MIL attention pooling setting, DPI scores 0.00 ROCAUC at all witness rates, indicating complete failure in that configuration.
Limitations
- Only two prior levels (high and low) for instance class were considered; richer, uncertainty-aware priors are left open.
- Evaluated on a semi-synthetic dataset with simulated noisy priors rather than real expert annotations.
- Applicable only to MIL-like settings with a separable attention module, limiting generality.
Experimental setup
Datasets: SARS-CoV-2 immune repertoires (Nolan et al., 2020) EBV-binding immune receptor sequences (Bagaev et al., 2020) MNIST-MIL (Ilse et al., 2018)
Benchmarks: semi-synthetic immune repertoire dataset MNIST-MIL
Metrics: bag-level ROCAUC instance-level PRAUC
Baselines: Vanilla DeepRC Average pooling (AP) DPI L1 LogReg L2 LogReg L2 SVC Conjunctive pooling (Conj)
Builds on: DeepRC ADMIL (attention-based deep MIL) Modern Hopfield networks
Closest papers by meaning
- Uncertainty Estimation by Flexible Evidential Deep Learning (NeurIPS 2025 · similarity 0.82)
- Are Multiple Instance Learning Algorithms Learnable for Instances? (NeurIPS 2024 · similarity 0.82)
- On Discriminative Probabilistic Modeling for Self-Supervised Representation Learning (ICLR 2025 · similarity 0.82)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.82)
- Componential Prompt-Knowledge Alignment for Domain Incremental Learning (ICML 2025 · similarity 0.82)
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · 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.