Error Correction Code Transformer
Authors: Yoni Choukroun, Lior Wolf
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate our method, we train the proposed architecture with three classes of linear block codes: Low-Density Parity Check (LDPC) codes [9], Polar codes [1] and Bose Chaudhuri Hocquenghem (BCH) codes [3]. All parity check matrices are taken from [11]. The proposed architecture is defined solely by the number of encoder layers N and the dimension of the embedding d. We compare our method to the BP algorithm, to the augmented hypernetwork BP algorithm of [21] (hyp BP), to the RNN architecture of [2] and to the very recent SOTA performance of [22] (AR BP). All the results were obtained from the corresponding papers. A comparison to the SCL Polar decoder [28] is given in Appendix D. The results are reported as bit error rates (BER) for different normalized SNR values (d B). We follow the testing benchmark of [21, 22]. During testing, our decoder decodes at least 105 random codewords, to obtain at least 500 frames with errors at each SNR value. |
| Researcher Affiliation | Academia | Yoni Choukroun Tel Aviv University choukroun.yoni@gmail.com Lior Wolf Tel Aviv University liorwolf@gmail.com |
| Pseudocode | Yes | Algorithm 1: Mask construction Pseudo Code |
| Open Source Code | Yes | Code available at https://github.com/yoni Lc/ECCT |
| Open Datasets | No | The paper mentions that "All parity check matrices are taken from [11]", which refers to a "Database of Channel Codes and ML Simulation Results". However, it does not provide concrete access information (e.g., specific dataset files, names of public benchmark datasets used for training, or formal citations with author/year for the specific data instances used for training) that would allow direct access to the actual training datasets. It refers to types of codes and a database of code properties rather than specific datasets with splits. |
| Dataset Splits | No | The paper specifies training parameters like "128 samples per minibatch" and "1000 epochs" but does not explicitly provide information regarding dataset splits for training, validation, and testing. It describes aspects of testing but does not detail a separate validation split or its size/composition. |
| Hardware Specification | Yes | Training and experiments have been performed on a 12GB Titan V GPU. |
| Software Dependencies | No | The paper mentions software components like "Adam optimizer", "GEGLU layers", "layer normalization", and "cosine decay scheduler" but does not provide specific version numbers for any libraries, frameworks (e.g., PyTorch, TensorFlow), or other software used for implementation or experimentation. |
| Experiment Setup | Yes | The Adam optimizer [14] is used with 128 samples per minibatch, for 1000 epochs, with 1000 minibatches per epoch. For N = 10 architectures we trained the models for 1500 epochs. ... The additive Gaussian noise is sampled randomly per batch in the {3, . . . , 7} normalized SNR (i.e. Eb/N0) range. ... We initialized the learning rate to 10 4 coupled with a cosine decay scheduler down to 5 10 7 at the end of the training. No warmup was employed [32]. |