Tiny Machine Learning QuickStart
ML that fits on tiny devices — constraints as the design surface.
Mechanisms
Fit inside memory, power, and energy budgets
Cloud habits (scale out) are the wrong reflex; the work is fit-to-budget.
Unlike cloud-based machine learning systems with virtually unlimited resources, TinyML development requires careful optimization to fit within tight memory constraints, limited processing power, and strict energy budgets.
Why it matters
Cloud habits (scale out) are the wrong reflex; the work is fit-to-budget.
Use when: Use when a team ports a cloud model and is shocked it will not fit.
Train heavy, deploy light
Asymmetry is the method: expensive training, cheap inference.
We can exploit this asymmetry by training our models on resource-heavy hardware, such as a desktop PC, and then converting them into a lightweight format that fits the constraints of our embedded hardware.
Why it matters
Asymmetry is the method: expensive training, cheap inference.
Use when: Use when training and deployment are treated as the same machine problem.
Arguments
Growth of models vs embedded limits
Progress-by-scale collides with devices that cannot scale.
Historically, the power of machine learning algorithms has increased over time, largely due to the growth in the size and complexity of the models.
Why it matters
Progress-by-scale collides with devices that cannot scale.
Use when: Use when a roadmap assumes bigger models are always better.
Evidence
Embedded hardware = limited CPU, program space, RAM
The envelope is the forcing function: fixed mass, fixed power, fixed link budget means every subsystem trades against every other — constraint is what turns a wish list into a design.
Due to size, efficiency, power, and cost constraints, embedded hardware is characterized by limited resources (CPU, program space, and RAM).
Why it matters
The envelope is the forcing function: fixed mass, fixed power, fixed link budget means every subsystem trades against every other — constraint is what turns a wish list into a design.
Use when: Use when someone argues "just use a bigger chip" as if cost and power were free.
Definitions
AI on the smallest, most constrained devices
The field is defined by scarcity, not by model cleverness alone.
This convergence has given birth to Tiny Machine Learning (often shortened as TinyML®, trademark of the EdgeAI foundation)—a field that brings the power of artificial intelligence to the smallest and most resource-constrained computing devices.
Why it matters
The field is defined by scarcity, not by model cleverness alone.
Use when: Use when "edge AI" is used as a buzzword without naming the constraint.