Score-based Source Separation with Applications to Digital Communication Signals
Authors: Tejas Jayashankar, Gary C.F. Lee, Alejandro Lancho, Amir Weiss, Yury Polyanskiy, Gregory Wornell
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results with RF mixtures demonstrate that our method results in a BER reduction of 95% over classical and existing learning-based methods. Our analysis demonstrates that our proposed method yields solutions that asymptotically approach the modes of an underlying discrete distribution. We demonstrate through experimental results (see 5) that α-RGS outperforms classical signal processing and annealed Langevin-dynamics-based approaches for RF source separation, with a 96% and 94.5% improvement in BER (averaged across different levels of interference) over traditional and existing learning-based methods, respectively. |
| Researcher Affiliation | Academia | 1Massachusetts Institute of Technology 2Universidad Carlos III de Madrid |
| Pseudocode | Yes | Algorithm 1 Proposed Method: α-posterior with Randomized Gaussian Smoothing (α-RGS) |
| Open Source Code | Yes | The code for reproducing our results can be found at https://github.com/tkj516/ score_based_source_separation and is also linked on our project webpage https:// alpha-rgs.github.io. |
| Open Datasets | Yes | We train diffusion models on different RF datasets i) synthetic QPSK signals with RRC pulse shaping (see 5.1), ii) synthetic OFDM signals (BPSK and QPSK) with structure similar to IEEE 802.11 Wi Fi signals; and iii) signals corresponding to Comm Signal2 from the RF Challenge [44], which contains datasets of over-the-air recorded signals. All synthetic datasets were created using the NVIDIA Sionna toolkit [45]. |
| Dataset Splits | Yes | All datasets contain between 150k – 500k samples and we use a 90-10 train-validation split during training. |
| Hardware Specification | Yes | We train all our models on 2 NVIDIA 3090 GPUs. Additionally, during data loading, we perform random time shifts and phase rotations on the OFDM (BPSK), OFDM (QPSK) and Comm Signal2 signals. ... Each set of separation experiments was conducted on a single NVIDIA V100 GPU. |
| Software Dependencies | No | The paper mentions using "NVIDIA Sionna toolkit [45]" for synthetic datasets and adopting the "Diffwave [24] architecture" for experiments, but it does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We use noise variance levels in the range (1e-4, 0.72) discretized into 50 levels. We train for 500k steps with early stopping on 2 x NVIDIA 3090 GPUs. Detailed hyperparameters for our training setup are provided in Appendix D. ... Our proposed algorithm uses ω = κ2 and is initialized with the MF solution given the mixture y (see 5.1). Note that κ can be equivalently described as the signal to interference ratio (SIR := 1/κ2 := Es s 2 2 /Eb κb 2 2 ). We assume that κ is known4 and use N = 20,000 with a cosine annealing learning rate schedule [46]. The OFDM mixtures use (ηmax, ηmin) = (5e-3, 1e-6) and the Comm Signal2 mixture uses (ηmax, ηmin) = (2e-3, 1e-6). |