Is it possible to write neural networks for microcontrollers or is it better to use FPGA?

Question

Python

I am engaged in robotics and study in the direction of Mechatronics. There is very little knowledge so far, there is experience with arduino. Googled, but would like to hear from people who may have experience and can offer advice. What will be more convenient and what is better for such a task? Is it possible to use Arduino or is it better to use MK separately and make your own boards? as an example: a self-learning robot that will be equipped with distance sensors and learn to distinguish objects by their size

Answer 1

The question is also whether it is necessary to write this for microcontrollers (inventing bicycles along the way) - or is it easier to put on board, for example, some arm with linux and tensorflow (for example) for similar functionality. And leave any realtime to microcontrollers, for example.
"self-learning robot"
One-shot reinforcement learning, or do we have unlimited bots? :-)

Answer 2

Let's go from the facts. A neural network is a model. The utility model is a very large model. Well, for example, a tensorflow model for recognizing a pair of Russian words (in conditions of different timbres and noise) will take something around 10 GB. Working with such a model will take a decent amount of resources. Where is the microcontroller?
For starters, it makes sense to figure it out and learn how to cook. Then the task of optimizing everything (storage and computation time) will already arise, so that it could somehow be stuffed into an FPGA or the like. Such things are definitely done, but this is a purely algorithmic / mathematical story.

Answer 3

I would make a system with a light brain on board, and leave everything heavy on the server. A good network interface will help.