Machine learning empowered communications and advanced signal processing

Key Research Areas & Activities

Neural receivers: By replacing individual components (channel estimation, equalization, and demapping) with a single deep neural network (DNN), we reduce the “error propagation” between blocks.
Semantic communications: By sifting research from sending bits to sending meaning, we increase the data, spectral, and energy efficiency.
Autoencoder-based Transceivers: Using autoencoders to jointly design the transmitter and receiver as a single optimized pair, specifically tailored for difficult channels like high-speed trains or underwater links.

AI-accelerators: Researching the integration of NPUs (Neural Processing Units) into the radio front-end to handle microsecond-level AI inference for tasks like interference cancellation. By employing AI accelerators, we can reduce the energy consumption of AI-native processes of the processing unit of the transceivers and/or the microcontroller.
Co-design software and hardware: Developing radio firmwares that aren’t “static.” Instead of a fixed algorithm for a filter, the radio uses a learned modelthat updates itself as the hardware ages or temperatures change. This is expected to minimize the transceives CO2 footprint and the response time, while maximizing security, resillience and reliability.
Energy-Efficient Inference: Quantizing (simplifying) complex deep learning models so they can run on low-power IoT radios without draining the battery.

Fingerprinting & Classification: Using Deep Learning to identify exactly what kind of signal is interfering (e.g., “that’s a microwave oven,” “that’s a military radar,” or “that’s a malicious jammer”) and adapting the wave shape to survive.
Collaborative Spectrum Sensing: Research into how multiple IDRs can share their “map” of the airwaves to create a city-wide, real-time database of spectrum availability.
Predictive Frequency Hopping: Using Reinforcement Learning to predict which frequencies will become congested in the next 500ms and proactively switching the radio to a clearer band.

Digital-to-Surface Mapping: Research on how the IDR software can directly control a metasurface antenna to create “holographic” beams that can track 100+ users simultaneously with zero moving parts.
Impairment Compensation: Using ML to “learn” the specific quirks of a radio’s own hardware (like a slightly noisy amplifier, in-phase and quadrature imbalance, and phase noise) and digitally canceling those errors before the signal is even sent.
Fluid Antennas: Exploring radios that use liquid metals or MEMS to physically reshape the antenna structure based on the AI’s recommendation for the current environment.

Deep Learning Beamforming: Training models to predict the best “beam” direction based on user location and environment history, without needing a full channel scan every ms.
CSI Compression: Using Computer Vision techniques (like CNNs) to compress Channel State Information (CSI) feedback. This allows the network to understand the environment without “clogging” the uplink with overhead data.
Predictive Handover: Using Recurrent Neural Networks (RNNs) or LSTMs to predict where a user is moving so the signal “follows” them before they even disconnect.

Dynamic Spectrum Access: ML agents that “listen” to the radio spectrum and autonomously find empty gaps to transmit data, avoiding interference in crowded 6G bands.
Slicing & Orchestration: Automatically “slicing” network resources (e.g., giving high priority to a remote surgery and low priority to a background download) in real-time.
Green Communications: DRL algorithms that learn when to put base stations into “sleep mode” to save energy without dropping calls.

Quantum Deep Reinforcement Learning (QDRL) for Slicing: Network slicing requires making thousands of micro-decisions per second about how to allocate bandwidth to different users. QDRL uses quantum circuits to speed up these decisions.
Quantum Combinatorial Optimization: Network orchestration is essentially a massive puzzle: “How do I route traffic through 10,000 nodes without a single bottleneck?” This is an “NP-Hard” problem where quantum computers excel.
Quantum-Enhanced Traffic Analysis & Security: QML can spot patterns in network traffic that are invisible to classical models, making it a key tool for 6G security.
Federated Quantum Learning (FQL): Since 6G is highly distributed, research is focused on how to train quantum models without moving sensitive user data to a central quantum hub.

Anti-Jamming via RL: Developing Reinforcement Learning agents that play a “cat and mouse” game with jammers, constantly evolving their modulation and coding to stay connected.
Protocol Obfuscation: Researching IDRs that can change their “language” (protocol) entirely on the fly to prevent an eavesdropper from even recognizing that a transmission is happening.
Federated Learning (FL): Training ML models across millions of user phones without the raw data ever leaving the device. Only the “learning” is sent to the tower.
Split Learning: Splitting a neural network between a low-power IoT device and a powerful Edge server to process complex signal tasks without draining the device’s battery.
Generative Adversarial Networks (GANs) for Data: Using GANs to create “synthetic” radio environments to train models when real-world data is scarce or sensitive.