SageMaker is an Amazon Web Service that enables customers to quickly and easily create machine learning models for cloud deployments.
It includes a set of standard machine learning algorithms such as k-means clustering, principal component analysis, neural topic modeling, and time series prediction. The context in which cloud-based machine learning models work is rarely static. Models often need to be updated as new training data becomes available or new use cases emerge.
However, simply retraining the model with the new data risks compromising the previously acquired knowledge of the model. Retraining the model with a combination of new and old data will work around this issue, but it can be very time consuming.
The design of the SageMaker system can help you to solve this deadlock. It also facilitates parallelization of model training and more efficiently optimizes the “hyperparameters” of the model, that is, the structural features of the model where variability can affect performance. For example, in neural networks, hyperparameters include characteristics such as the number of network layers, the number of nodes per layer, and the learning rate of the network. Optimal settings for model hyperparameters vary from task to task, and adjusting hyperparameters for a particular task is usually a difficult process of trial and error.
Our system design addresses these issues by distinguishing between models and model states. In this context, the state is a summary of the data the model has seen so far.
As a simple example, if your model calculates a moving average of an incoming stream of numbers. The state of the model includes both the sum of all the numbers displayed and their quantity. If the model saves this state, when a new stream of numbers arrives next week, you can continue to increase both values without having to read the numbers you have already seen again.
Of course, most machine learning models perform more complex tasks than simple averaging, and the information that the state needs to capture varies from task to task. For example, you can include a representative sample of the data displayed. We used SageMaker to identify individual state variables for each machine learning algorithm we support.
This means that the cost of training a model, both in terms of time and system resources, will change in proportion to the amount of new training data. If the training time is scaled hyperlinearly, and with a sufficient amount of data, the training may expire and fail. The averaging example shows another aspect of the system. Must work with streaming data. This means that you can view each training example only once and stop the series of examples at any time. At each such breakpoint, you should be able to integrate what you have learned to create a working modern model.
