Image recognition on Arm Cortex-M with CMSIS-NN

Yes, it is possible to run TensorFlow models on Arm Cortex-M by using TensorFlow Lite Micro. Click here for more information.

CMSIS-NN provides optimized functions, but whether the weights are hard coded at compile time vs. provided at run time depends on the use case. For example, OpenMV cam uses CMSIS-NN back end with weights provided at run time. The examples provided in CMSIS GitHub repo use weights that are hard coded during binary compilation. 

CMSIS-NN has all the optimized functions required to run MobileNet model. The main challenge is whether you have enough memory along with the Arm Cortex-M core to fit the MobileNet weights and layer outputs.

Just a standard MCU, no dedicated hardware.

The CIFAR-10 object recognition model requires 77KB for the weights that are stored in Flash memory and 40KB for the activation's that get stored in RAM.

This is a very good question - given a range of microcontrollers and the flexibility of neural network sizes, how do you pick the right one? I recommend first training a network of any size to solve your problem to the accuracy you require. Then begin simplifying that network whilst maintaining the desired accuracy. Finally you can compare memory and FLOP counts to get an approximate sizing before benchmarking on hardware. A good example of this in practice is our keyword search paper: Machine Learning on Arm Cortex-M Microcontrollers.

It depends on your requirements in terms of latency, accuracy and throughput. You can do some level of motion recognition on Arm Cortex-M4, and you can expect to run larger, more accurate versions of those networks on Arm Cortex-M7.

 Yes, Mbed OS is not required. 

CMSIS-NN is an inference library and does not support training.

All Arm Cortex-M processors support CMSIS-NN. The ones with SIMD will have an overall better performance.