The life cycle of machine learning is defined as a cyclical process. There are three phases in this cyclical process – pipeline development, training, and inference. These phases are possible from the data acquired by the data scientist and data engineers who develop, train, and serve machine learning models using the large data that are involved in various applications.
It is an iterative process to build a machine learning model. This means that for a successful deployment towards the end of the project cycle, the steps in the model need to be repeated multiple times. Even after deployment, this model needs to be maintained and adapted to a changing environment. So, the process of developing, deploying, and managing a machine learning model for a specific application is called the machine learning lifecycle. In the lifecycle, there are several steps.
The entire process begins by determining the business objective of implementing such a machine learning model. The business objective for a bank can be to decrease the number of fraudulent transactions from the total transactions that take place at the bank. Then, data collection and exploration come into the picture. After determining the business objective for implementing the machine learning model, data is collected for the machine learning task. Then comes the exploratory data analysis and data visualizations to understand what the data is saying, and which processes are needed to make the data ready before model training.
The data must satisfy the business objective. Data is then ready for the model or made ready for the machine learning model. To make the data ready for the machine learning model, data is cleaned, split into training, validated, and put through test sets, and feature engineering. Feature engineering means transforming data to better represent the business objective. Automated feature training becomes possible with AutoML tools.
The prepared data is then used to train the developed machine learning model. This is an iterative process. During this process, several different algorithms are tested, the suitable model is selected and finetuned to its hyperparameters to achieve the best performance. Hyperparameters means model parameters that influence the learning process that are not interpreted through the data. An example of this is the size of a neural network.
The machine learning model is then evaluated on a test set so that it performs correctly for the specific business model and replicates data on use cases. Before deployment, all issues are addressed. These include excessive resource requirements and insufficient performance. In the former, the model might require too much memory and take too long to process information. So, the model performance is optimized by software engineers. In the latter, the cost of deploying the model can be more than its benefits. The model might not give or might not identify accurate predictions from itself. All predictions might then need to be verified by a human, making false positives expensive. The model might not have high accuracy rates.
Finally, the model after being fine-tuned, is deployed to make the predictions. There are several ways the model can be deployed. In online deployment, the model is deployed via an API to respond to requests in real-time and serve predictions. In batch deployment, the model is integrated into a batch prediction system. In embedded model, the model is embedded in an edge or mobile device.
After deployment, the model is then monitored so that its performance is up to the mark. Unless the previously mentioned machine learning model to catch bank frauds is being regularly updated, it might not be able to catch new fraud types. Then it becomes sort of redundant. For models that are trained in specific intervals, a new iteration of the development process can be launched.
Challenges of machine learning models lifecycle management
There is a lot of manual labour involved in every step and transition between the steps. It means that data scientists need to sit down and collect, analyze, and process the data for each application by themselves. They need to fine-tune the models by observing the older models, comparing them to new ones and developing new ones. To prevent performance degradation, a lot of time is allocated in projects to model management and all of it is manual.
Data scientists can build machine learning models by themselves. However, around 55% of businesses working on machine learning models have not deployed a model to production. This is because data scientists have to collaborate with software engineers, designers, business professionals and others to build a machine learning model. Hence, the process becomes much more complicated when so many cooks are involved.
However, as more and more machine learning models are deployed, it becomes challenging to manually manage the entire process. It may require data scientists to split into teams and take up responsibilities like development, management, monitoring, etc. For models that are trained in specific intervals, a new iteration of the development process can be launched.