Reproduce “MinGRU-Based Encoder for Turbo Autoencoder Frameworks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rick Fritschek, Rafael F. Schaefer
Researchers from Technische Universität Dresden introduced a new encoder architecture for the Turbo Autoencoder framework, integrating a minimal Gated Recurrent Unit (minGRU) with a Mamba block. This design enables parallel computation for sequence processing, achieving error-rate performance comparable to existing CNN-based Turbo Autoencoders for a rate 0.5 code (k=64, n=128) while demonstrating faster training convergence.
Key insights
- Modern, efficient RNN variants like minGRU, when combined with structured state-space models like Mamba, can overcome the historical computational inefficiencies and scalability limitations of RNNs in channel coding.
- The simplified minGRU design significantly reduces parameters and allows for parallel processing, addressing the linear time dependency of traditional RNNs for long sequences.
- Integrating a Mamba block's multiplicative gating mechanism around the minGRU enhances training dynamics and overall performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.