This Week's Top AI & Tech Innovations: Federated Learning, AI Processors, and More

JD Launches Fedlearn Federated Learning Platform

JD Digital Technology Group has unveiled its self‑developed federated learning platform, Fedlearn, designed to meet data‑privacy and regulatory requirements while enabling AI systems to collaboratively train on distributed data. Fedlearn features three key aspects: it exchanges only intermediate values instead of raw data or model parameters, introduces a centralized data exchange concept that decouples data transfer from participants, and adopts an asynchronous computation framework that significantly speeds up model training.

Facebook Uses AI to Accelerate Renewable‑Energy Catalyst Discovery

Facebook, in collaboration with Carnegie Mellon University, is applying artificial‑intelligence techniques to discover new electrocatalysts that can convert solar and wind energy into alternative fuels. The partnership claims AI models can expand the searchable catalyst space to millions or billions of candidates, dramatically speeding up the identification of promising materials.

ARM China Announces "Zhouyi" Z2 AIPU AI Processor

ARM China has released the Z2 AIPU, an AI‑specific processing unit that doubles single‑core performance to 4 TOPS compared with its predecessor and reduces chip area by 30%. The processor supports up to 32 cores for a total of 128 TOPS, targeting high‑end security cameras, intelligent cockpits, ADAS, and edge servers. Architectural optimizations maintain the same micro‑architecture while improving efficiency.

NVIDIA Introduces BlueField DPU for Data‑Center Tasks

NVIDIA has launched its BlueField data processing unit (DPU), which moves many data‑center infrastructure functions onto a single chip. Integrated into a SmartNIC, the DPU uses standard PCIe interconnects and is already being shipped to early customers as the first‑generation BlueField‑2, with broader availability expected in 2021.

Chrome Begins Deploying HTTP/3 with IETF QUIC

The Chromium team announced that Chrome is rolling out support for HTTP/3, which is built on the QUIC protocol. Early measurements show QUIC reduces Google Search latency by over 2 %, cuts YouTube rebuffering time by more than 9 %, and improves throughput on both PC and mobile clients by several percent.

Canonical Says Microsoft Won’t Port Windows to Linux

Amid speculation about Microsoft’s Linux integration strategy, Canonical desktop team member Hayden Barnes refuted claims that Microsoft plans to rebuild Windows on the Linux kernel. While some observers, such as Eric S. Raymond, have suggested such a move, Barnes emphasizes that these are personal opinions and not official plans.

Academic News: New Scenarios for Stereo Matching

This article reviews recent conference work on stereo matching, focusing on three innovations: 360° panoramic image disparity estimation, binary/quantized disparity representations for resource‑constrained deployment, and the creation of synthetic binocular datasets by combining monocular depth estimates with existing data. These approaches open new research directions beyond traditional follow‑up studies.

Academic News: IndRNN‑Based Mobile Sensor Action Recognition

Researchers from the University of Electronic Science and Technology of China and Shandong University applied the IndRNN architecture—longer and deeper than conventional RNNs—to mobile sensor data for action recognition. The model achieved strong performance, especially when combined with transfer learning and model‑fusion post‑processing, earning a runner‑up prize at the UbiComp SHL Challenge 2020.