Dilated CNNs for Periodic Signal Processing: A Low-Complexity Approach
New model matches AR methods with 1 observation and minimal compute.
A new paper from researchers Eli Gildish, Michael Grebshtein, and Igor Makienko introduces R-DCNN, a computationally efficient deep learning method for denoising periodic signals and estimating waveforms. Published on arXiv (2604.21651), the work targets core signal processing tasks in speech, music, medical diagnostics, radio, and sonar. While deep learning has improved over classical techniques, it typically requires heavy compute and per-observation retraining. R-DCNN breaks that mold by using a dilated convolutional neural network combined with a lightweight resampling step that aligns signals with different fundamental frequencies, allowing a single network to generalize without retraining.
The method achieves performance comparable to state-of-the-art classical approaches like autoregressive-based techniques and conventional DCNNs trained individually for each observation. Crucially, it operates under strict power and resource constraints, making it ideal for IoT devices, wearables, and edge computing. The paper includes 16 pages and 8 figures, demonstrating how R-DCNN balances efficiency and accuracy. This opens the door for real-time, low-power signal processing in environments where computational budgets are tight, without sacrificing denoising or estimation quality.
- R-DCNN trains on a single observation and reuses weights across varying frequencies via resampling.
- Matches performance of autoregressive methods and individually-trained DCNNs with far lower compute.
- Designed for resource-constrained environments like IoT, wearables, and edge devices.
Why It Matters
Enables high-accuracy signal denoising on low-power devices, expanding AI's reach to edge computing.