Tiered On-Device Sensing / 分层端侧感知

Definition

Tiered on-device sensing is a low-power architecture where small sensors and chips continuously monitor for simple signals, then wake larger local or cloud compute only when a scene requires richer perception, recording, or reasoning.

Current Synthesis

The RayNeo source uses always-on camera as the concrete case. 李鸿伟 says a small camera-side CNN can sample frames at low frequency, a simpler chip can recognize notable scenes, and a larger chip can wake for video or complex processing only when needed.

The concept matters because AI glasses have less battery, space, thermal headroom, and social tolerance than phones. It links Edge-Cloud AI Boundary to product fit: without local filtering and wake-up logic, First-Person AI Memory / 第一视角AI记忆 becomes either too power-hungry, too cloud-dependent, or too privacy-invasive for daily wear.

Key Claims

• Always-on wearable AI requires local filtering before heavy model calls.
• The architecture separates low-frequency perception, simple scene detection, and expensive recording or reasoning.
• Power, heat, and battery constraints make sensing hierarchy a product requirement, not only an engineering optimization.
• Trigger design shapes privacy and data minimization because not every moment should become stored or cloud-processed data.
• The concept is source-scoped to RayNeo’s described architecture until supported by additional implementations.

Evidence

• Three-layer scheme - No.221 雷鸟 CEO：新技术越来越多，我们为什么还需要一副智能眼镜？ describes camera-side CNN sampling, simple-chip scene recognition, and larger-chip escalation.
• Battery constraint - No.221 雷鸟 CEO：新技术越来越多，我们为什么还需要一副智能眼镜？ presents the design as a way to support Live Log despite small glasses batteries.
• Chip logic - No.221 雷鸟 CEO：新技术越来越多，我们为什么还需要一副智能眼镜？ later links glasses chips to extreme low power, small/large core design, RTOS, and possible low-power islands.

Counterevidence & Qualifications

The source gives an architectural description but not measured battery life, false-positive rates, privacy safeguards, or external verification of RayNeo’s implementation.

What Changed

• Created the concept from the No.221 RayNeo interview.

Related Concepts

• Edge-Cloud AI Boundary - broader split between local and cloud AI work.
• On-Device AI - local execution layer used by tiered sensing.
• First-Person AI Memory / 第一视角AI记忆 - memory use case that depends on sensing hierarchy.
• AI Glasses Product Fit / AI眼镜产品适配 - product-fit question shaped by battery and latency.
• AI Hardware Privacy Exchange / AI硬件隐私交换 - privacy tradeoff affected by local filtering.