Neural Joint Source-Channel Coding
Authors: Kristy Choi, Kedar Tatwawadi, Aditya Grover, Tsachy Weissman, Stefano Ermon
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, we test NECST on several grayscale and RGB image datasets, obtaining improvements over industry-standard compression (e.g, Web P (Google, 2015)) and error-correcting codes (e.g., low density parity check codes). |
| Researcher Affiliation | Academia | Kristy Choi 1 Kedar Tatwawadi 2 Aditya Grover 1 Tsachy Weissman 2 Stefano Ermon 1 1Department of Computer Science, Stanford University 2Department of Electrical Engineering, Stanford University. |
| Pseudocode | No | The paper describes its methods and training procedures in detail, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | We provide reference implementations in Tensorflow (Abadi et al., 2016), and the codebase for this work is open-sourced at https://github.com/ermongroup/necst. |
| Open Datasets | Yes | We experiment on randomly generated length-100 bitstrings, MNIST (Le Cun (1998)), Omniglot (Lake et al. (2015)), Street View Housing Numbers (SVHN) (Netzer et al. (2011)), CIFAR10 (Krizhevsky (2009)), and Celeb A (Liu et al. (2015)) datasets |
| Dataset Splits | No | The paper mentions using training and test sets but does not explicitly provide specific details about validation set splits or methodology. |
| Hardware Specification | No | The paper mentions experiments were run on 'CPU' and 'GPU' and notes speedup differences, but does not provide specific hardware details such as CPU or GPU models, memory, or other detailed specifications. |
| Software Dependencies | No | The paper states that implementations are in 'TensorFlow' but does not provide specific version numbers for TensorFlow or any other software dependencies. |
| Experiment Setup | Yes | Empirically, we found that using that using a continuous relaxation of the discrete latent features led to worse performance at test time when the codes were forced to be discrete. Therefore, we used VIMCO (Mnih & Rezende (2016)), a multi-sample variational lower bound objective for obtaining low-variance gradients. VIMCO constructs leave-one-out control variates using its samples, as opposed to the single-sample objective NVIL (Mnih & Gregor (2014)) which requires learning additional baselines during training. Thus, we used the 5-sample VIMCO objective in subsequent experiments for the optimization procedure, leading us to our final multi-sample (K = 5) objective: |