NL-CALIC Soft Decoding Using Strict Constrained Wide-Activated Recurrent Residual Network.

Area of Science:

• Digital Communications
• Signal Processing
• Machine Learning

Background:

• Soft-decoding of near-lossless codes faces challenges in strict error boundary enforcement, adaptability to compression ratios, and implementation efficiency.
• Existing methods often lack end-to-end solutions or are not optimized for hardware acceleration.

Purpose of the Study:

• To develop an effective and efficient soft-decoding strategy for near-lossless codes.
• To address the challenges of peak absolute error (PAE) boundary, variable quantization steps, and hardware implementation.

Main Methods:

• Proposed a Wide-activated Recurrent structure with a normalized Tanh activation strategy (WRSD).
• Designed a normalized Tanh activation strategy for strict PAE boundary control.
• Developed a lightweight, wide-activate recurrent structure suitable for GPU/FPGA implementation.

Main Results:

• WRSD outperforms state-of-the-art soft decoders.
• Achieved superior performance with less than 5% of the parameters of existing methods.
• Each computation node requires minimal storage (<64KB), facilitating caching on consumer GPUs.

Conclusions:

• The proposed WRSD technique effectively solves key challenges in soft-decoding of near-lossless codes.
• WRSD offers a computationally efficient and hardware-friendly solution for near-lossless data compression.
• The method demonstrates significant improvements in performance and resource utilization.