Xenofon's Feedback

[jasonjabbour]

Chapter 2 ML Systems:

• AI vs ML, Collect all quasi-synonymous terms (TinyML, Edge ML, Edge AI, Embedded AI, Embedded ML, Embedded Intelligence, …) to help readers navigate through the literature
• Cloud ML: Geopolitical concerns, reliance on foreign powers, reduced sovereignty
• Edge ML: improve definition to include/exclude telecoms towers. For me the real difference is whether ML runs on computers controlled by one organisation or outsourced?

Chapter 3 DL Primer

• Traditional ML primer? Agree with main thesis, perhaps the book would be richer if TML is detailed a bit more, e.g. with list of algorithms, pros cons among TML algorithms, and learning resources

Chapter 4 AI Workflow

• Traditional vs Embedded AI: Highlight that the development and execution of software often occurs on different machines, requiring cross-compilation, flashing, etc, but also making testing/debugging more challenging

Chapter 5 Data Engineering

• 5.3 Data Sourcing: Foundation: Sensors, ADCs, Sampling Frequency/Resolution, Multimodal sensing, time synchronization
• 5.4 Data Storage: Anonymization, specific rules derived from ethics approval committees or local laws (GDPR)

Chapter 8 Efficient AI

• Efficient Numbers: Highlight that sub-8bit quantization needs specialized hardware to be leveraged. A generic CPU may just pad zeros.
• Efficient Numerics -> Basics: it reads as if integers cannot also be 16, 32, and 64 bits, perhaps it should be rephrased to highlight why a float16 is preferable than an int16
• Data? Full-System Optimization?