MLOps refers to the adoption of a DevOps approach to the management and oversight of machine learning (ML). Its aim is to automate the processing of the tasks involved in a machine learning application and, in so doing, accelerate them.

The first task of every MLOps team is to determine the general sequence and classification of essentially ML task the team will undertake. Once they have done this, the team will likely discover their sequence can be divided into four phases which constitute the MLOps lifecycle:

• Setup and establishment of the data infrastructure that will supply inputs to the ML models, and the hosting of data to which the output of those models will pertain
• Delegation of roles and responsibilities for the management and oversight of source code repositories, implementation of the virtual infrastructure required for hosting ML workspaces, the pipelines for continuous implementation and deployment (CI/CD), and the monitoring and observation of active models for performance and reliability
• The workflow for the training, analysis, and evaluation of ML models, often referred to by MLOps professionals as the inner loop. This includes the procedures for data ingestion during training, experimentation and testing, and the final registry of a working ML model for production
• The deployment and maintenance of ML models in production, referred to as the outer loop. This phase incorporates the introduction of ML models into the CI/CD pipeline, continual evaluation for performance, and analysis of operating characteristics for possible retraining

In an era where so many of us are actually chatting with AI on a regular basis now, we need to remember that machine learning isn’t entirely about language.

At its heart, the point of all machine learning is to produce a model that can act as a mathematical function. Given a large set, or vector, of input values, this function is designed to produce a result that represents a prediction. Training this function produces a model that progressively improves at rendering more accurate predictions.

Training the ML function well means enabling it to determine the most informative aspects of the input data, so that its function can produce the most accurate prediction from the minimum number of inputs. This is feature extraction — finding the features that communicate the relevant aspects of the data to the function, even if no one can accurately relate what those features mean.

The whole point of MLOps is to make the operations that constitute the machine learning process regular and repeatable, to such an extent that they may become automated. Feature extraction is one of the most important “inner-loop” ML processes that can be automated. Automating machine learning does not mean leaving it unattended. It means making it manageable by skilled operators. If the steps involved in training and operating an ML model are regular, trainable, and repeatable, MLOps professionals can be relied upon to make predictive analysis capabilities available and ready throughout their organizations.

Relevant links

• AmpereOne: A Game Changer for Efficient, Scalable Compute, and AI Workloads
• Wallaroo and Ampere Accelerate AI Inference by 7X
• Lower Your Costs with Ampere and Arrow’s Integrated Solutions
• AmpereOne: Cores are the new MHz by Jeff Geerling