List-based and Trellis-based Decoding of the oFEC code
Description
This project investigates the creation of patterns in product code decoding. Alternatively, low-complexity Trellis-based approaches are considered.
Supervisor:
Error Correction for Binary Neural Networks in Noisy Environments
Binary Neural Networks, Channel Coding, Lattices, Robust Inference, Crossbar, Memristors
Description
This thesis investigates coding-theoretic methods to increase the robustness of binary neural networks (BNNs) implemented on noisy hardware platforms, such as memristor crossbars or radiation-prone memories used in space applications. The objective is to design, model, and evaluate decoding strategies and coding schemes ensuring reliable neural computation when both weights and activations are binary and hardware noise significantly affects performance. The work builds upon recent results [1] showing how channel coding principles can be applied to BNN layers.
[1] E. Dupraz and F. Leduc-Primeau, "Channel Coding for Binary Neural Networks Implemented with Noisy Memristor Crossbars," IEEE ISTC, 2025.
For detailed information, please refer to the attached PDF.
Prerequisites
- Solid background in digital communications, information and coding theory.
- Understanding of neural network fundamentals.
- Programming experience in Python (PyTorch/TensorFlow).
Recommended lectures:
Channel Codes for Iterative Decoding and Machine Learning for Communications
Contact
andreas.strasshofer@tum.de
riccardo.schiavone@dlr.de