Anyone familiar with AI research, or at least those that have done some reading on the history in the throws of hype will be familiar with AI “summers” and “winters”. These are terms that describe the historic, cyclical periods of alternating enthusiasm and technology breakthroughs, followed by disillusionment and reduced interest in the field of AI research and development.
In the previous post, we looked at the components of Machine Learning Ops at a high level and briefly touched on each stage of the life cycle. With this in mind, we can now look at the finer processes of a machine learning infrastructure solution and begin to explore the tooling and technologies used in their curation.
Looking at the image below, you’ll notice that the machine learning code is only a small portion of the overall solution. The bulk of tooling within machine learning infrastructure exists to support the capture, integration and monitoring of data ingested and produced by the model. The quality and accuracy of model produced data is largely dependant on the quality of data ingested as well as the measures which have been put in place to either enrich or verify the data. The smallest of changes in a data set can compound over time and ultimately lead to quite a sizeable, negative impact to model performance. Most of us will be familiar with the term GIGO ( Garbage In Garbage Out ) and this remains an important concept to keep in consideration when building machine learning solutions.
As the tidal wave of AI continues to surge with companies such as OpenAI releasing more sophisticated iterations of their products, practices such as MLOps are becoming all the more relevant in the modern landscape of cloud infrastructure.
Inspired by DevOps, Machine Learning Ops or Deep Learning Ops is the practice and methodologies surrounding the creation of machine learning models, the iteration and training of these models in addition to the curation of automatable workflows that allow these models to be deployed or utilised within Production environments.
This post aims to give an overview on what MLOps is and the key areas involved in understanding how we create an automatable, scalable workflow for the consistent delivery of performant models.