Research map / Continual learning / Online and replay-based
Neurocoder: General-Purpose Computation Using Stored Neural Programs
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 Neurocoder to recode the classifier layers | MLP, LeNet, and DenseNet · CIFAR10 and CIFAR100 | accuracy: 52.06 -> 54.86; 23.31 -> 26.24; 75.71 -> 78.92; 42.73 -> 47.21; 93.61 -> 95.61; 68.37 -> 72.20 | Helped |
| Add Neurocoder to the MLP for incremental task learning | Split MNIST · Split MNIST incremental task | accuracy: 93.75$\pm$3.28 -> 96.54$\pm$1.39; 98.02$\pm$0.89 -> 99.01$\pm$0.19; 98.14$\pm$0.43 -> 98.35$\pm$0.74; 98.69$\pm$0.20 -> | Helped |
| Add Neurocoder to the MLP for incremental domain learning | Split MNIST · Split MNIST incremental domain | accuracy: 53.55$\pm$1.27 -> 58.46$\pm$0.46; 57.83$\pm$2.74 -> 62.28$\pm$4.03; 64.37$\pm$2.40 -> 69.89$\pm$1.72; 64.41$\pm$3.36 -> | Helped |
| Use recurrent multi-step attention instead of single-step attention (ablation) | Neurocoder with GRU · Polynomial auto-regression | mean squared error | Helped |
| Add Neurocoder to the GRU output layer | GRU · Polynomial auto-regression | mean squared error | Helped |
| Add Neurocoder to A3C actor and critic networks | A3C · Frostbite and Montezuma’s Revenge | game score: 1,500 to 3,000; almost zero | Helped |
| Increase the orthogonal-loss weight (ablation) | Neurocoder-augmented ResNet · Split CIFAR10 | accuracy | Helped |
| Use a larger program memory (ablation) | Neurocoder-augmented ResNet · Split CIFAR10 and CIFAR100 | accuracy: 15%; 10% | Helped |
| Combine Neurocoder with Synaptic Intelligence | ResNet · Split CIFAR10 | accuracy: above 80% | Helped |
| Combine Neurocoder with Synaptic Intelligence | ResNet · Split CIFAR100 | accuracy: 10% | Helped |
| Use NSM conditional computation instead of the original MLP | MLP · Split MNIST incremental task | accuracy: 93.75 -> 87.55; 98.02 -> 96.63; 98.14 -> 91.44; 98.69 -> 98.87; 97.00 -> 93.94; 98.23 -> 96.11 | Hurt |
Closest papers by meaning
- Neural Stored-program Memory (ICLR 2020 · similarity 0.88)
- Data-Efficient Learning with Neural Programs (NeurIPS 2024 · similarity 0.86)
- Differentiable Programs with Neural Libraries (ICML 2017 · similarity 0.86)
- Neuro-Symbolic Program Synthesis (ICLR 2017 · similarity 0.86)
- HOUDINI: Lifelong Learning as Program Synthesis (NeurIPS 2018 · similarity 0.86)
- Encodings for Prediction-based Neural Architecture Search (ICML 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.