You are cordially invited to the faculty candidate seminar which will be given by Dr. Emre Özfatura this Friday on December 27, 2024 at 9:30 a.m.

The short information about the seminar, its summary and the biography of Dr. Özfatura are given below:

Title: “AttentionCodes: AI-aided iterative error correction codes:”
Place: Electrical and Electronics Engineering  Seminar Room
Date and Time: Friday, December 27, 2024 / 9:30 a.m.

Deep neural network (DNN)-assisted channel coding, such as low-complexity neural decoders for existing structured codes, or end-to-end neural-network-based auto-encoder designs have received significant interest in recent years due to their improved performance and flexibility. Such designs are particularly attractive for channels for which high-performing structured codes do not exist, such as the feedback channel or multi-user networks. In the scope of our research, we consider communication with an active feedback link, where the receiver can encode its feedback over the noisy feedback channel, which combines the challenges of feedback and multi-user problems. We propose novel encoding and decoding algorithms for both the transmitter and the receiver, built upon the transformer architecture, self-attention mechanism in particular. In the proposed solution, we combine the self-attention mechanism with an active feedback strategy in order to further improve the block error rate (BLER) performance and to reduce the amount of feedback required to make the overall design more attractive for practical implementations. We further propose a variable length coding strategy, which inspired from semi-supervised learning, and seek a better trade-off between the BLER performance and the communication budget, finally we have also show that such strategy is not limited to point-to-point communication scenario but can be generalized to multi-access scenario as well.