There are a few possible approaches to deploying a ML model to a microcontroller.
The main limiting factor to deployment on microcontollers is that ML models are usually a representation of a set of parameters that are intended to be used as input to a prediction algorithm alongside a new datapoint. Most such models assume the presence of an accompanying library that implements the algorithm in question. However, a microcrontoller may use an exotic chip architecture, or have very several or unusual resource constraints that prevent these standard libraries from being deployed easily.
Presumably you will already have some way to get input into your microcontroller and to program it in order to call some function that you can write. If not, you will need to first figure out how to do that, and the right methods depend on your microcontroller. A common approach is to write assembly code or code in a very limited subset of C or another language. An alternative is to find a distribution of an interpreter for another language (e.g. Java, Python) that has been compiled to work on your chip. Either way, you will need some way to program the chip.
Presuming you can program the chip, you have two fundamental challenges in deploying the model:
Most models are trained with very wide floating point numbers for their parameters. For example, 128-bit numbers may be used. On a standard computing environment, the CPU or GPU will be equipped to perform operations on wide datatypes efficiently. On a microcontroller, you may be limited to 8-bit or 16-bit integers. To work with your model parameters in an environment like this, you will need to either make the parameters smaller (usually by rounding them to fit in a much smaller numeric format, a process called "quantization"), or by finding or writing software that can simulate the operations you want (probably addition and multiplication) on a large datatype that is represented as a collection of smaller datatypes. The first approach may make the model perform poorly. The second may make model prediction very slow.
You need an implementation of the algorithm. Some algorithms like linear regression, linear discriminant analysis, or even decision trees are extremely easy to implement prediction for, and may require only addition, multiplication, and/or comparison. You might be able to write these yourself in a simple subset of C, or even in assembly (for example, prediction with linear regression should be just a simple loop). Other algorithms, like deep neural networks, may contain more complex operations, and may contain many such operations performed in complicated sequences. For these, you generally will need to find an distribution of a library that implements the algorithms, or compile one yourself. Compiling one yourself will require setting up a build toolchain for your specific microcontroller, and can be quite involved.
