Research map / Optimization / Optimizers
HOUDINI: Lifelong Learning as Program Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 typed program synthesis with gradient-trained neural modules rather than standalone or evolutionary approaches | Houdini on graph shortest-path tasks · GS1 shortest_path_street | RMSE: 1.88 vs 4.96 (Standalone); 8.32 (Houdini EA) | Helped |
| Transfer the pretrained shortest-path relaxation function to the new street-image domain | Houdini · GS2 shortest_path_street | RMSE: 3.48 vs 6.92 | Helped |
| Reuse the learned counting network for a new digit class | Houdini · CS2 Task 3 count_digit | RMSE: 0.40 -> 0.34 | Helped |
| Reuse a high-level counting network across the MNIST-to-NORB toy-image domain shift | Houdini · CS3 Task 3 count_toy | RMSE | Helped |
| Use a fold combinator to sum a list of digit representations | Houdini · SS Task 2 sum_digits | RMSE: 2.15 | Helped |
| Enable type-directed pruning instead of generating programs without types (ablation) | Houdini program search · CS1 Task 1, program size 4 | number of programs generated: 8182 -> 2 | Helped |
| Enable type-directed pruning instead of generating programs without types (ablation) | Houdini program search · CS1 Task 4, program size 6 | number of programs generated: 6474938 -> 175 | Helped |
| Use top-down synthesis rather than evolutionary synthesis (ablation) | Houdini · Lifelong task sequences | test error | Helped |
| Use 10 rather than 9 graph-convolution iterations (ablation) | Houdini · GS2 Task 2 shortest_path_mnist | RMSE: 1.73 -> 1.57 | Helped |
| Use 10 rather than 9 graph-convolution iterations (ablation) | Houdini · GS1 Task 2 shortest_path_street | RMSE: 2.02 -> 1.88 | Helped |
Closest papers by meaning
- From Perception to Programs: Regularize, Overparameterize, and Amortize (ICML 2023 · similarity 0.89)
- Differentiable Programs with Neural Libraries (ICML 2017 · similarity 0.88)
- Improving the Universality and Learnability of Neural Programmer-Interpreters with Combinator Abstraction (ICLR 2018 · similarity 0.87)
- Lifelong Learning with Weighted Majority Votes (NeurIPS 2016 · similarity 0.86)
- Separating the 'what' and 'how' of compositional computation to enable reuse and continual learning (NeurIPS 2025 · similarity 0.85)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.85)
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.