Research map / Label-efficient and robust learning / Recognition and long-tail
Towards Automated Knowledge Integration From Human-Interpretable Representations
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 informed meta-learning, which conditions a meta-learned prior on human-interpretable knowledge (e.g., natural language), and instantiates it as Informed Neural Processes evaluated on synthetic functions, weather temperature, and CUB few-shot classification.
Problem
Incorporating prior knowledge in natural-language or other native formats into ML models automatically, rather than through manual inductive-bias engineering, especially in low-data and noisy settings.
Approach
Extends meta-learning to be knowledge-conditioned so that the learned prior p(f|K) maps knowledge representations to hypothesis-space priors; instantiated as Informed Neural Processes that add a knowledge encoder fused with data representations via a latent variable, with theory showing knowledge cannot increase expected predictive KL in expectation.
Key results
- On weather prediction, knowledge gives 0-shot log-likelihood improvements of 57.7% (setup A, min/max vector) and 48.1% (setup B, natural-language forecast) over uninformed predictions.
- On CUB 5-way 0-shot, INP (A) reaches 87.5% accuracy and INP (B) 87.4%, versus no zero-shot baseline for NP; at 1-shot 5-way, INP (B) gets 89.1% vs NP 82.2%.
- On synthetic sinusoidal regression, informing the INP with two of three parameters yields larger log-likelihood gains than one, and INP without knowledge performs on par with NP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition predictions on class attributes (INP A) instead of using only context examples (NP) | INP on CUB-200-2011 · 5-way 1-shot | accuracy (%): 82.2 -> 88.1 | Helped |
| Condition predictions on class textual descriptions (INP B) instead of using only context examples (NP) | INP on CUB-200-2011 · 5-way 1-shot | accuracy (%): 82.2 -> 89.1 | Helped |
| Condition predictions on generated class descriptions (INP C) instead of using only context examples (NP) | INP on CUB-200-2011 · 5-way 1-shot | accuracy (%): 82.2 -> 85.1 | Helped |
| Condition predictions on knowledge in place of context-only NP | INP on CUB-200-2011 · 10-way 1-shot | accuracy (%): 73.3 -> 82.2; 73.3 -> 81.9; 73.3 -> 77.0 | Mixed |
| Provide class-specific knowledge (A, B, or C) to the INP | INP on CUB-200-2011 · 5-way and 10-way classification, 1, 3, 5, and 10-shot | accuracy (%) | Mixed |
| Provide two values of daily minimum and maximum temperature as knowledge | INP on USCRN temperature prediction · zero-shot | log-likelihood improvement: 57.7% | Helped |
| Provide a GPT-4-generated natural-language weather forecast as knowledge | INP on USCRN temperature prediction · zero-shot | log-likelihood improvement: 48.1% | Helped |
| Provide knowledge about the shifted parameter b | INP on synthetic sinusoidal regression · OOD test tasks with b ~ N(3, 1) | log-likelihood | Helped |
| Omit knowledge at test time | INP on synthetic sinusoidal regression · synthetic test tasks | log-likelihood | No clear effect |
| Provide knowledge about one or two function parameters | INP on synthetic sinusoidal regression · varying number of context points | relative ΔAUC | Helped |
| Use knowledge about parameter b rather than no knowledge | INP on synthetic sinusoidal regression · epistemic uncertainty across x in [-2, 2] | epistemic uncertainty | Helped |
Surprising findings
- Informed meta-learners can underperform uninformed meta-learners when the number of meta-training tasks is too small relative to knowledge complexity.
- Zero-shot INP on CUB with GPT-4 generated per-class descriptions (setup C) reaches only 50.3% at 5-way, much lower than setups A or B.
Limitations
- Requires a meta-training set of related tasks with paired knowledge, which may be hard to obtain.
- Neural approximation does not strictly enforce knowledge constraints (e.g., min/max bounds can be exceeded).
- Informed meta-learning cannot guarantee correctness of knowledge interpretation unlike model-based methods.
Experimental setup
Datasets: CUB-200-2011 U.S. Climate Reference Network sub-hourly temperature synthetic sinusoidal functions
Benchmarks: CUB-200-2011 U.S. Climate Reference Network synthetic sinusoidal regression
Metrics: log-likelihood classification accuracy delta AUC epistemic uncertainty
Baselines: Neural Process (NP) MAML prototypical networks
Builds on: Neural Processes MAML Conditional Neural Processes Attentive Neural Processes meta-learning with meta-data
Closest papers by meaning
- Sharing Knowledge for Meta-learning with Feature Descriptions (NeurIPS 2022 · similarity 0.88)
- MARS: Meta-learning as Score Matching in the Function Space (ICLR 2023 · similarity 0.87)
- Human-like Category Learning by Injecting Ecological Priors from Large Language Models into Neural Networks (ICML 2024 · similarity 0.85)
- More Flexible PAC-Bayesian Meta-Learning by Learning Learning Algorithms (ICML 2024 · similarity 0.85)
- Provable Meta-Learning with Low-Rank Adaptations (NeurIPS 2025 · similarity 0.85)
- Context-Aware Meta-Learning (ICLR 2024 · similarity 0.84)
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.