Research map / Architectures and efficiency / RNNs and normalization
Towards Scaling Difference Target Propagation by Learning Backprop Targets
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 |
|---|---|---|---|
| Train feedback weights with L-DRL instead of DRL | DTP feedback pathway · CIFAR-10 LeNet, output layer alignment | angle between feedforward and feedback weights: 18.7° to ≈3° | Helped |
| Train feedback weights with L-DRL instead of DRL | DTP · LeNet, CIFAR-10 batch; feedforward update alignment | angle to BP updates: ≈90° to ≲35°; DRL ≲79° | Helped |
| Use DTP with L-DRL feedback training rather than s-DDTP | LeNet · MNIST test | accuracy (%): 98.59_{\pm 0.16} -> 98.93_{\pm 0.04} | Helped |
| Use DTP with L-DRL feedback training rather than s-DDTP | LeNet · Fashion-MNIST test | accuracy (%): 88.86_{\pm 0.44} -> 90.91_{\pm 0.17} | Helped |
| Use DTP with L-DRL feedback training rather than s-DDTP | LeNet · CIFAR-10 test | accuracy (%): 76.33_{\pm 0.27} -> 85.33_{\pm 0.32} | Helped |
| Replace s-DDTP with p-DDTP, training feedback and feedforward weights together without pretraining or extra feedback-only epochs (ablation) | DDTP · MNIST test | accuracy (%): 98.59_{\pm 0.16} -> 98.58_{\pm 0.13} | No clear effect |
| Replace s-DDTP with p-DDTP, training feedback and feedforward weights together without pretraining or extra feedback-only epochs (ablation) | DDTP · Fashion-MNIST test | accuracy (%): 88.86_{\pm 0.44} -> 89.42_{\pm 0.69} | Helped |
| Replace s-DDTP with p-DDTP, training feedback and feedforward weights together without pretraining or extra feedback-only epochs (ablation) | DDTP · CIFAR-10 test | accuracy (%): 76.33_{\pm 0.27} -> 72.15_{\pm 0.29} | Hurt |
| Use DTP with L-DRL instead of p-DDTP with DRL | LeNet · CIFAR-10 test | accuracy (%): 72.15_{\pm 0.29} -> 85.33_{\pm 0.32} | Helped |
| Use DTP with L-DRL feedback training | VGG-like · CIFAR-10 | accuracy (%): BP 89.07_{\pm 0.22} -> DTP 89.38_{\pm 0.20 } | Helped |
| Use DTP with L-DRL feedback training | VGG-like · ImageNet 32×32 validation | Top-5 accuracy: 61.25 -> 60.54 | Hurt |
Closest papers by meaning
- Efficient Target Propagation by Deriving Analytical Solution (AAAI 2024 · similarity 0.93)
- Forward Learning with Top-Down Feedback: Empirical and Analytical Characterization (ICLR 2024 · similarity 0.87)
- Overcoming Rank Collapse in Feedback Alignment (NeurIPS 2026 · similarity 0.86)
- Improving equilibrium propagation without weight symmetry through Jacobian homeostasis (ICLR 2024 · similarity 0.86)
- Two Tales of Single-Phase Contrastive Hebbian Learning (ICML 2024 · similarity 0.85)
- Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation (ICML 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.