What is machine learning?
Machine learning is the science that provides a computer the ability to learn without being explicitly programmed. It’s based on the usage of algorithms that analyze data, learn from it and make a specified prediction. Running machine learning models often requires a lot of computing power and it is common to run the algorithms in the cloud. Edge computing brings that capability to the devices.
In the context of AI@Edge the machine learning models are run in containers that are deployed to the devices using Azure IoT Hub.
Creating an AI model
To get started you’ll need to:
- Decide which machine learning framework (and neural network) you are going to use
- Select tools to be used
- Understand the high level process
AI models, especially ones for vision that transform a camera input into labels and objects, use often neural networks. Neural network is a series of highly connected algorithms that mimic the way human brains operate to recognize relationships and pattern in a data set. Neural networks are used to cluster and classify data.
Deep neural networks (DNN) are neural networks with additional complexity in terms of having multiple layers between inputs and outputs. SqueezeNet is an example of DNN.
Machine Learning Frameworks
There are different training frameworks used by the data science community. These frameworks are used to build, i.e. train, deep learning models using their libraries with the data (training data). The framework libraries are used to represent the details of the underlying machine learning algorithms in the model. Examples of training frameworks are Tensorflow, Pytorch and Caffe2.
These frameworks also have inference engines that are used to execute the deep neural network models. The inference engine is also referred to as the runtime. This runtime is capable of translating the operators in the DNN model into computation steps that are executed when input data is provided to the model.
Selecting a neural network and machine learning framework
When creating a complex AI model being able to select a correct neural network and ML framework is essential. This is generally a complex area and we are targeting to provide additional information and guidance in the near future. Especially for vision there the frameworks may have different performance and accuracy with certain use cases like facial recognition, anomaly detection, activity recognition. The underlying complexity also often makes the AI model hardware dependent. Microsoft, among other companies, is solving this problem with ONNX.