Research map / Label-efficient and robust learning / Recognition and long-tail
Differentiable Programs with Neural Libraries
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 a differentiable program interpreter with a shared, trainable neural-function library instead of a purely neural model | Lifelong Add2x2 and Apply2x2 tasks · Add2x2 and Apply2x2 | accuracy | Helped |
| Use a differentiable program interpreter with a shared neural-function library instead of MTNN-2 | Lifelong Add2x2 tasks · Add2x2 | accuracy on earlier tasks | Helped |
| Continue training the shared neural functions on later tasks | NTPT · Add2x2:top | accuracy | Helped |
| Train on only 2-digit expressions and use a program with loop-based control flow | NTPT · Math, expressions up to 16 digits (31 symbols) | generalization performance | Helped |
| Replace NTPT with an LSTM trained on examples up to 5 digits | Math · Math, expressions up to 16 digits (31 symbols) | generalization performance | Hurt |
| Train a purely neural model on each Add2x2 task independently with only 1k examples | NTPT and purely neural baselines · Add2x2 | accuracy | Hurt |
| Train a purely neural model on individual tasks with at least 4k examples | Baseline models · Add2x2 | accuracy | Helped |
| Switch the dominant training task from Add2x2 to Apply2x2 | MTNN-2 shared perceptual network · Add2x2 tasks during lifelong training | accuracy | Hurt |
| Use progressive neural columns with earlier-task weights frozen | Progressive Neural Network · Lifelong Add2x2 and Apply2x2 | catastrophic forgetting | Helped |
| Use a tuned task-specific architecture rather than a shared-depth neural baseline | MTNN-2 versus PNN · Lifelong Add2x2 and Apply2x2 | overall performance | No clear effect |
| Use free loop structure in the differentiable interpreter | NTPT Math model · Math, finding a correct program | convergence: 2/100 | Hurt |
Closest papers by meaning
- Learning Differentiable Programs with Admissible Neural Heuristics (NeurIPS 2020 · similarity 0.89)
- Data-Efficient Learning with Neural Programs (NeurIPS 2024 · similarity 0.85)
- Breaking Neural Network Scaling Laws with Modularity (ICLR 2025 · similarity 0.84)
- LILO: Learning Interpretable Libraries by Compressing and Documenting Code (ICLR 2024 · similarity 0.84)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.84)
- Large Language Models are Interpretable Learners (ICLR 2025 · 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.