This blog summarises recent works by the academic team on the work package – Learning to Compress and Communicate. The focus is on emerging techniques in semantic communication, image compression, point cloud transmission, channel state information compression, and hybrid joint source-channel coding. These works reflect the growing trend towards task-oriented compression, neural implicit representations, and adaptive strategies for efficient data delivery in bandwidth-constrained and error-prone wireless environments.
Compression Beyond Pixels: Semantic Compression With Multimodal Foundation Models (Wu, Gunduz et al) This paper introduces a task-agnostic semantic compression framework leveraging multimodal foundation models like CLIP. It employs a Product Quantization Variational Autoencoder (PQVAE) with a shared codebook to compress embeddings at extremely low bitrates (~2–3×10⁻³ bpp). The approach preserves zero-shot robustness across diverse tasks and datasets, outperforming traditional image compression methods while using less than 5% of their bitrate. Applications include edge intelligence and privacy-preserving vision tasks.
LotteryCodec: Searching the Implicit Representation in a Random Network (Wu, Gunduz, Dragotti et al) proposes the Lottery Codec Hypothesis, demonstrating that untrained subnetworks within randomly initialized networks can serve as synthesis networks for image compression. The method compresses an image into a binary mask and latent modulations, avoiding transmission of full network weights. It achieves state-of-the-art rate-distortion performance among overfitted codecs and even surpasses VTM in BD-rate, while offering adaptive decoding complexity via adjustable mask ratios.
Over-the-Air Learning-based Geometry Point Cloud Transmission (Gunduz et al) This work addresses wireless delivery of 3D point clouds through three frameworks: SEPT for small-scale point clouds using SNR-adaptive modules; OTA-NeRF, which represents point clouds as neural radiance fields and transmits model weights; and OTA-MetaNeRF, a meta-learning approach that sends lightweight latent vectors instead of full models. These methods outperform traditional compression and digital schemes, eliminate cliff effects, and enable real-time transmission under fading channels.
MIMO Channel as a Neural Function—Implicit Neural Representations for Extreme CSI Compression (Wu, Gunduz et al) The CSI-INR framework reimagines channel state information feedback as a neural implicit representation. It uses a meta-learned base network shared across devices and modulation-based feedback with quantization and entropy coding for extreme compression. Results show up to 11 dB NMSE improvement and 5× lower bit rates compared to benchmarks, supporting flexible feedback strategies for massive MIMO systems.
A Deep Joint Source-Channel Coding Scheme for Hybrid Mobile Multi-hop Networks (Gunduz et al) DJSCC combines analog DeepJSCC for the first wireless hop with neural compression for subsequent wired hops, mitigating noise accumulation and avoiding cliff effects. Enhanced with SNR-adaptive and rate-adaptive modules, it dynamically adjusts to channel conditions and rate-distortion trade-offs. The framework outperforms digital baselines in PSNR and SSIM, reduces storage by 15×, and processes images in under 0.03s, making it suitable for edge-to-cloud scenarios.
Future Directions
Potential research avenues include:
- – Integrating semantic and structural compression for multi-modal data streams.
- – Integrating semantic and structural compression for multi-modal data streams.
- – Combining LotteryCodec principles with semantic embeddings for hybrid codecs.
- – Extending OTA frameworks to multi-modal AR/VR content and dynamic environments.
- – Scaling CSI-INR for cooperative and multi-user MIMO scenarios.
- – Deploying h-DJSCC in real-world edge-to-cloud testbeds and exploring cross-layer optimization.
These directions aim to unify efficiency, adaptability, and robustness across diverse communication and compression tasks.
*Hub acknowledges the use of Microsoft CoPilot to assist with the drafting of this blog.
