On‑Device AI and Federated Learning: Era Background, Theory, and Practical Applications

Background and Era Overview

From 1G (1986) to the emerging 6G era, communication technologies have evolved from analog voice to massive IoT connectivity, leading to an explosion of devices and data.

Third AI Wave

The current AI wave is powered by three pillars: massive data, solid theoretical foundations, and ever‑increasing compute power. IDC predicts 175 ZB of new data per year by 2025; AI‑related publications exceeded 30 k per year in 2020; and training compute demand now doubles every three to four months.

Federated Learning

Proposed by Google in 2016, federated learning enables collaborative model training without sharing raw data, addressing privacy and data‑island challenges.

Horizontal FL: same features, different samples.

Vertical FL: same samples, different features.

Federated Transfer Learning: finds commonalities across feature‑sample spaces.

On‑Device AI

Edge AI moves inference to the device, reducing latency, protecting privacy, lowering bandwidth and inference costs, and allowing personalized models.

Typical Models

Decision Trees (including Random Forest, GBDT) – simple, interpretable, suitable for small data and low‑resource devices.

Neural Networks – high‑capacity models requiring more data and hardware acceleration, used for tasks such as image/video super‑resolution.

Practical Cases

Google keyboard: gradient aggregation on phones.

Video pre‑loading optimization using on‑device feature reporting.

Dynamic super‑resolution for video streams.

Local feed‑ranking and autonomous driving perception pipelines.

Benefits

Edge AI with federated learning reduces bandwidth and server load while improving CTR, GMV, UV, and PV metrics.