Research map / Label-efficient and robust learning / Recognition and long-tail
Controlling Neural Networks with Rule Representations
Sungyong Seo, Sercan O Arik, Jinsung Yoon, Xiang Zhang, Kihyuk Sohn, Tomas Pfister
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeepCTRL with tuned alpha instead of fixed-lambda Task & Rule regularization | DeepCTRL · Double pendulum test | rule verification ratio (%): 71.6 -> 80.2 | Improved |
| Apply DeepCTRL instead of TaskOnly on weaker-correlation Neg0.1 items (model from Neg0.2) | DeepCTRL · Sales Neg0.1 | MAE: 0.05 -> 0.045 | Improved |
| Transfer model from source Neg0.2 to Neg0.1 target (TaskOnly degradation) | TaskOnly · Sales Neg0.1 | MAE: 0.041 -> 0.05 | Hurt |
| Transfer model from source Neg0.2 to Neg0.1 target (DeepCTRL degradation) | DeepCTRL · Sales Neg0.1 | MAE: 0.042 -> 0.045 | Hurt |
| Sample alpha from Beta(1,1) (uniform) instead of Beta(0.1,0.1) (ablation) | DeepCTRL · Double pendulum | MAE and verification ratio | Hurt |
| Vary Beta(beta,beta) prior around beta=0.1 (ablation) | DeepCTRL · All datasets | MAE | No clear effect |
| Replace stochastic coupling with Concat or Add of rule and data representations (ablation) | DeepCTRL · Double pendulum | MAE at intermediate alpha | Hurt |
| Concatenate alpha to input instead of two-passage architecture (ablation) | DeepCTRL · Double pendulum | task error and verification ratio | No clear effect |
| Adapt rho every epoch instead of fixing it from initial losses (ablation) | DeepCTRL · Double pendulum | MAE and verification ratio | No clear effect |
| Increase perturbation upper bound u (ablation) | DeepCTRL · Cardiovascular | cross entropy / accuracy at alpha>0 | Hurt |
| Add rule via perturbation loss, compare TaskOnly vs DeepCTRL on Source cardiovascular data | DeepCTRL · Cardiovascular Source | cross entropy | Hurt |
| Increase alpha up to 1.4 (extrapolation beyond training range) | DeepCTRL · Cardiovascular Target 1 | cross entropy: 1.0 -> 1.4 | Improved |
Closest papers by meaning
- Deep Neural Networks Constrained by Decision Rules (AAAI 2019 · similarity 0.84)
- Learning Accurate and Interpretable Decision Rule Sets from Neural Networks (AAAI 2021 · similarity 0.84)
- Self-explaining deep models with logic rule reasoning (NeurIPS 2022 · similarity 0.84)
- Neural Compositional Rule Learning for Knowledge Graph Reasoning (ICLR 2023 · similarity 0.84)
- HyperLogic: Enhancing Diversity and Accuracy in Rule Learning with HyperNets (NeurIPS 2024 · similarity 0.82)
- Neuro-Symbolic Temporal Point Processes (ICML 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.
