Research map / Learning theory and algorithms / Prediction and decision losses
Deepcode: Feedback Codes via Deep Learning
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 nonlinear RNN instead of a linear feedback scheme | Feedback code · AWGN feedback, rate 1/3, 1dB (σ_F²=0.01) | BER: 7.83e-04 -> 1.56e-04 | Helped |
| Use a nonlinear RNN instead of a linear RNN (ablation) | Bit-by-bit feedback code · AWGN feedback, rate 1/3, 1dB (σ_F²=0.01) | BER: 0.0046 -> 1.56e-04 | Helped |
| Add zero padding to the information sequence (ablation) | RNN(tanh) feedback code · AWGN feedback, rate 1/3 | BER | Helped |
| Add trainable power allocation across raw and coded bit streams (ablation) | RNN(tanh)+ZP · AWGN feedback, rate 1/3 | BER: about an order of magnitude | Helped |
| Add trainable bit-position-specific power allocation (ablation) | RNN(tanh)+ZP+W · AWGN feedback, rate 1/3 | BER | Helped |
| Replace raw feedback with RNN-encoded feedback | Deepcode · AWGN with noisy, delayed feedback | BER | Helped |
| Allow K-step delayed feedback (ablation) | Deepcode · AWGN with noisy feedback | BER | No clear effect |
| Concatenate Deepcode as an inner code with a turbo outer code | Deepcode · AWGN, rate 1/9, increasing block length | BER | Helped |
| Use an unrolled RNN at longer block lengths (ablation) | Deepcode · AWGN with noiseless feedback, block lengths 50 to 500 | BER: 50 to 500 | No clear effect |
| Use random initialization instead of initialization from a pretrained non-feedback code (ablation) | Jointly trained encoder and decoder · AWGN feedback | BER | Helped |
| Use active coded feedback rather than passive feedback | Delayed-feedback neural code · AWGN with noisy feedback | BER | Helped |
| Use bit-by-bit linear RNN instead of jointly encoding bits with RNN memory (ablation) | Feedback code · AWGN feedback, rate 1/3, 1dB (σ_F²=0.01) | BER: 1.56e-04 -> 0.0046 | Hurt |
Closest papers by meaning
- Robust Non-Linear Feedback Coding via Power-Constrained Deep Learning (ICML 2023 · similarity 0.90)
- Communication Algorithms via Deep Learning (ICLR 2018 · similarity 0.87)
- Turbo Autoencoder: Deep learning based channel codes for point-to-point communication channels (NeurIPS 2019 · similarity 0.85)
- KO codes: inventing nonlinear encoding and decoding for reliable wireless communication via deep-learning (ICML 2021 · similarity 0.84)
- DeepPolar: Inventing Nonlinear Large-Kernel Polar Codes via Deep Learning (ICML 2024 · similarity 0.83)
- Universal Sample Coding (NeurIPS 2024 · similarity 0.80)
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.