Developing Edge AI Solutions

Fraunhofer-Gesellschaft at a glance:

• Application-oriented research for economy and industry
• 30,000+ employees, 76 institutes, 3.0 billion € financial volume
• Over 70% from industrial/public-funded research, ~30% basic federal funding

AI hierarchy at Fraunhofer IIS (from smallest to largest):

Tiny ML → Edge ML → Efficient ML → ML → AI

Edge AI vs. Cloud AI comparison:

4 advantages of Edge AI (#fast, #energy-efficient, #private, #tiny):

1. Processing on energy-efficient sensor nodes

2. Lowest latency = fast response

3. Private data stays on device

4. Fits microcontrollers, reuses existing hardware

Edge AI Fast Track — 4 steps (pipeline):

1. Data — annotation with VLM (Visual Language Models) + SAM (Segment Anything Model)

2. Choose model — type selected by expert; architecture found by NAS (Neural Architecture Search); parameters by AutoML

3. Compress — "Deep Compression": Pruning (remove less important weights) + Quantization (lower numerical precision)

4. Deploy — integration + quality assurance in the field

Model compression techniques:

• Pruning: removes less important neurons/connections
• Quantization: uses lower numerical precision (e.g., Float32 → UInt8)
• Example: MicroYolo compressed from 8 MB → 0.8 MB via pruning + quantization

Unsupervised Anomaly Detection:

• Defects are rare and appear in unforeseen ways → severe data imbalance → supervised approaches impractical
• Training: learn a model of "normality" from only non-defective samples
• Inference: any significant deviation from normality → flagged as anomalous

TinyML application clusters:

• Industrial IoT (condition monitoring, predictive maintenance)
• Environment (earthquake detection)
• Healthcare (fitness tracker, health monitor)
• Agriculture (animal & plant monitoring)

Embedded system characteristics for Edge AI (exam example question topic!):

• Small size ✓
• Low weight ✓
• Low energy requirement ✓
• Real-time capability ✓
• MHz range processor ✓ (not GHz like desktop CPUs)
• Does NOT require continuous internet connection ✓
• Does NOT have high computing power (by design)