In this module, you will explore some of the key concepts in machine learning, such as feature extraction and classification models, in the context of signal processing. You will understand the importance of training and evaluation in the machine learning workflow, and the constraints involved when using microcontrollers for this. At the end of the module, you will complete a practical lab exercise, to implement some simple machine learning models for activity recognition, using accelerometer data. To do so, you will be shown how to use Anaconda and Python to work with datasets.

This module dives deeper into a powerful and widely used model in Machine Learning: the artificial neural network. These can analyze large quantities of input data in complex ways, in order to solve classification problems, such as identifying objects in an image. In order to run neural networks on small microprocessors, these models need to be as streamlined as possible. So you will also look at the complexity of a typical neural network, and see some techniques to reduce this complexity, such as quantization. In the lab, you will continue building a classifier for activity recognition, but this time using a neural network on an Arm STM32 microprocessor. For this, you will be introduced to the TensorFlow Python library, which is also popular for many applications in machine learning.

Neural networks can be used to solve complex classification problems, as you have already seen. In this module, you’ll discover a more advanced model: the convolutional neural network. These are important for image processing, as they can interpret relationships between adjacent pixels, but they are also used in other applications such as financial modeling. This is a new and modern technique so you’ll be learning about the cutting edge of machine learning, and the recent trends in this field. In the lab, you’ll develop a convolutional neural network for audio processing, and optimize it for both accuracy and performance. This would allow it to give good results on a small device without draining the battery or delaying the response.

The algorithms used in modern machine learning can be very complex, and require many iterations of innovation and testing by computer scientists. This is especially true for the optimized algorithms required by microprocessors! Thankfully, you do not need to implement these algorithms yourself, as they are available in libraries, such as CMSIS-NN, developed by Arm. This module shows you how this library can be used for machine learning—for example for image processing using convolutional neural networks. In the lab exercise, you also have the opportunity to use CMSIS-NN to develop a simple model for the CIFAR-10 dataset, using CUBE AI.

For machine learning to perform well, even on the smallest devices, it is essential to optimize the models to minimize their memory footprint and the number of operations required to perform inference tasks. In practice, this allows portable devices to be more responsive, and extends their battery life. In this last module, you’ll explore some of the cutting-edge techniques used to optimize neural networks, such as using fixed-point arithmetic in place of floating-point arithmetic. To consolidate your learning, you will develop the best machine learning model that you can, that would be able to run on an ArmCortex-M microprocessor, using a toolkit such as CMSIS-NN.