Periodic Reporting for period 2 - RENEW (Reinventing Energy Efficiency in Communication Networks)
Période du rapport: 2022-12-01 au 2024-05-31
To this day, communications engineering has closely followed the seminal guidelines developed by Claude E. Shannon in 1948, which were mostly influenced by the telephone network of those days. The widespread use of mobile communications and the advent of machine-to-machine communications nowadays entail an exponential increase in data rates and the available models used during the system design are no longer sufficient to design power-efficient, low-latency, high-speed communication systems. The overarching aim of RENEW is to further increase the data rates of the global telecommunication network while, at the same time, addressing its non-negligible environmental impact. By fundamentally revisiting the transceiver processing algorithms of the core parts of the communication network, RENEW has the potential to overcome the limitations of current design methodologies and to significantly reduce the complexity and energy consumption of the network. Capitalising on cutting-edge results in the fields of machine learning, reinforcement learning, optimisation techniques and neuromorphic computing, RENEW will reinvent the design of communication transmitters and receivers by introducing sparsely connected blocks that realise highly parallelisable transceivers guaranteeing high throughputs with low energy consumption. RENEW will explore novel concepts for extremely energy efficient receivers based on spiking neural networks, promising efficiency gains by multiple orders of magnitude. The viability of the RENEW concepts will be demonstrated in applications of high relevance such as high-speed optical communication networks or low-power IoT applications. The RENEW concept has the potential to yield novel energy efficient communication systems.
The research explored several innovative approaches to enhance communication systems across multiple objectives, focusing on the design of low-complexity channel codes, development of advanced transceivers using atomic neural blocks (ANBs), and the application of spiking neural networks for communications.
Objective 1 highlighted efforts in channel code design, yielding mixed results but notable successes with product codes optimized for hybrid decoding, significantly reducing energy consumption while maintaining high performance. These codes are poised to play a crucial role in future optical communication systems requiring low energy use. Further advancements include a pruning technique for quantum LDPC codes, addressing current challenges of latency and power consumption in quantum computing communications.
Objective 2 detailed the creation of ANB-based transceivers through both bottom-up and top-down approaches. The research led to the development of superior phase estimators with simpler configurations and adapted modulation formats, showcasing improved performance and reduced complexity. The work also revisited the factor graph framework, optimizing ANBs for equalization and symbol detection tasks, achieving near-optimal performance with significantly less complexity. A novel structural optimization method was introduced to find the most efficient ANB configurations, balancing energy use and complexity.
Objective 3 investigated reinforcement learning for energy-efficient resource allocation in NOMA and classical wireless communications.
Objective 5 focused on utilizing spiking neural networks for communications, examining how communication signals can be represented through spikes for potential use on neuromorphic hardware. This exploration resulted in identifying an equalizer as a key component for implementation using spiking neural networks, achieving comparable performance to classical DFE equalizers with significantly improved energy efficiency. The optimized equalizer for optical IM/DD transmission systems further enhanced energy savings by reducing spike density.
In summary, the research achieved significant advancements in communication system design, from developing energy-efficient coding schemes and transceivers to pioneering the use of spiking neural networks in communications. These contributions are expected to significantly impact future optical communication systems, quantum computing, and the implementation of neuromorphic hardware, marking a substantial step forward in the field.
Objective 1 highlighted efforts in channel code design, yielding mixed results but notable successes with product codes optimized for hybrid decoding, significantly reducing energy consumption while maintaining high performance. These codes are poised to play a crucial role in future optical communication systems requiring low energy use. Further advancements include a pruning technique for quantum LDPC codes, addressing current challenges of latency and power consumption in quantum computing communications.
Objective 2 detailed the creation of ANB-based transceivers through both bottom-up and top-down approaches. The research led to the development of superior phase estimators with simpler configurations and adapted modulation formats, showcasing improved performance and reduced complexity. The work also revisited the factor graph framework, optimizing ANBs for equalization and symbol detection tasks, achieving near-optimal performance with significantly less complexity. A novel structural optimization method was introduced to find the most efficient ANB configurations, balancing energy use and complexity.
Objective 3 investigated reinforcement learning for energy-efficient resource allocation in NOMA and classical wireless communications.
Objective 5 focused on utilizing spiking neural networks for communications, examining how communication signals can be represented through spikes for potential use on neuromorphic hardware. This exploration resulted in identifying an equalizer as a key component for implementation using spiking neural networks, achieving comparable performance to classical DFE equalizers with significantly improved energy efficiency. The optimized equalizer for optical IM/DD transmission systems further enhanced energy savings by reducing spike density.
In summary, the research achieved significant advancements in communication system design, from developing energy-efficient coding schemes and transceivers to pioneering the use of spiking neural networks in communications. These contributions are expected to significantly impact future optical communication systems, quantum computing, and the implementation of neuromorphic hardware, marking a substantial step forward in the field.
In several areas, we went significantly beyond the state of the art. In particular, we have designed new low-complexity channel codes, which we expect to further improve until the end of the project.
Furthermore, we have managed to investigate the first atomic neural blocks using both model-based algorithmic learning and the factor graph framework. We have proposed a novel structural optimization algorithm for factor-graph based communication receivers.- We expect to converge both methods until the end of the project to have a holistic atomic-block-based receiver.
Finally, we have proposed first receiver components implemented using spiking neural networks and we expect to have a larger library of transceiver components built using spiking neural networks until the end of the project.
Furthermore, we have managed to investigate the first atomic neural blocks using both model-based algorithmic learning and the factor graph framework. We have proposed a novel structural optimization algorithm for factor-graph based communication receivers.- We expect to converge both methods until the end of the project to have a holistic atomic-block-based receiver.
Finally, we have proposed first receiver components implemented using spiking neural networks and we expect to have a larger library of transceiver components built using spiking neural networks until the end of the project.