The Thirty-Ninth Annual Conference on Neural Information Processing Systems
INFORMED AI researcher Haotian Wu presented papers at this prestigious conference, highlighting a shift in how AI can be used to reconceptualize transmission and low-complexity compression. The presented papers were:
Realism and Fidelity: Two Sides of a Coin in Deep Joint Source-Channel Coding. Haotian Wu, Weichen Wang, Di You, Pier Luigi Dragotti, Deniz Gündüz (Imperial College London)
MoRIC: A Modular Region-based Implicit Codec for Image Compression. Gen Li, Haotian Wu, and Deniz Gündüz (Imperial College London)
The specific theories from these papers are:
- W²-DeepJSCC for adaptive image transmission balancing realism and fidelity,
-
MoRIC for modular, region-based low complexity compression,
Why these developments matter for adaptive AI compression the future of Edge AI
As AI moves from cloud servers to edge devices—think AR glasses, autonomous drones, and smart cameras—the challenge isn’t just making models smarter. It’s making them computation efficient, perceptually aligned, and bandwidth-aware. In real-world networks, bandwidth fluctuates, devices have limited compute, and humans care about what looks right—not just what’s pixel-perfect. Haotian’s two contributions at NeurIPS 2025 in San Diego tackle these issues head-on.
These innovations aren’t just academic—they’re blueprints for next-generation AI streaming, semantic communication systems, and AR/VR experiences that feel natural even under network constraints.
W²-DeepJSCC: Balancing Realism and Fidelity in Wireless Image Transmission
Background. Wireless image delivery often faces a trade-off: pixel fidelity vs. visual realism. Under poor signal conditions, traditional metrics like PSNR fail to capture what humans actually perceive.
W²-DeepJSCC introduces a channel-adaptive neural codec that dynamically shifts between perceptual realism and high-fidelity reconstruction based on signal-to-noise ratio (SNR).
- Saliency-Guided Perception–Fidelity Adapter (SG-PFA): Prioritizes perceptually important regions when bandwidth is tight.
-
Wavelet Wasserstein Distortion (WA-WD): A new perceptual metric inspired by human vision for better alignment with human preference, enabling fine-grained control over textures and details without the heavy compute of diffusion models.
W²-DeepJSCC enables:
- Human-centric optimization: Aligns image quality with what people actually see.
- Bandwidth-aware streaming: Perfect for AR/VR, telemedicine, and remote inspections.
- Low complexity: Outperforms diffusion-based methods in perceptual quality while remaining lightweight for edge devices.
MoRIC: Modular Region-Based Image Compression for Smarter Streaming
Background. Traditional codecs and even learned neural codecs treat an image as one big block. That’s inefficient when some regions matter more than others—like faces in a video call or instruments in a medical scan.
Solution : MoRIC (Modular Region-based Implicit Codec) introduces a divide-and-conquer strategy for image compression. Instead of one monolithic model, MoRIC splits an image into semantic regions (e.g., foreground objects vs. background) and assigns lightweight neural networks to each. It then uses adaptive contour coding to transmit region boundaries efficiently and a global modulation network to keep everything visually coherent.
What MoRIC can do:
- Fine-grained control: Compress important regions first, refine later.
- Layered compression: Send a quick preview, then progressively sharpen details as bandwidth allows.
- Efficiency: MoRIC achieves state-of-the-art rate–distortion performance with ultra-low decoding complexity, making it ideal for surveillance, mobile streaming, and IoT cameras.
- Scalability: Supports semantic-aware layered compression, paving the way for AR/VR streaming and privacy-preserving transmission.
The Big Picture: Smarter, Leaner, Human-Aligned AI
Both papers share a common philosophy: optimize for perception, modularity, and communication frugality. MoRIC shows how region-aware compression can make streaming scalable and privacy-conscious. W²-DeepJSCC proves that realism and fidelity aren’t mutually exclusive—they can coexist in a single adaptive model.
Together, these advances point to a future where:
- AR/VR feels seamless, even on spotty networks.
-
Healthcare imaging reaches clinicians at low bit cost and faster decoding without sacrificing diagnostic quality.
-
Smart cities and robotics coordinate efficiently under large data volume exchange without flooding networks.”
