How are your models changing?
Many are those entities that have a lifetime, living things, inert ones, and nowadays, digital bodies. For the first ones, we know a lot about when they are born, how they grow, and when they die. For the second ones, even if they don’t have a biological lifetime, inert things emerge from something and they can also be transformed to become something else. As for the third ones, when are the digital entities born, and how do they experience transformation?
Talking about the lifetime of a digital entity is quite similar to talking about it for inert things. Let’s think about a rock. There’s a point for the formation of a rock starting from magma. Then, there are many ways in which a rock can change. It can be back to the magma point after melting, or it can be disintegrated by many other types of transformations. If geology is too much, let’s think about another tool, like a simple pencil. We gather some materials, like wood and polishers, and we build a pencil. After being in the hands of others, pencils are consumed, or broken and even divided into small pieces to make tiny pencils. Our original pencil has lived a life. As a matter of fact, this process is not different for digital or virtual entities, such as our machine learning models.
A model is born
These digital tools are born when we build, train, fit, or estimate our model. This phase pretty much starts with having an analytical objective, data, computers, algorithms, and everything else data scientists are by now well aware of. Whatever other tools you gather, never forget about the analytical or scientific objective so your final model means something and serves a specific need. When is your model born? When you finish training it and save it for employment/deployment, the lifetime of this tool has begun.

What’s ahead of this newborn? That will depend on the analytical objective, so that’s why we cannot forget that part when building it. This model may serve prediction tasks, interpretation of indicators, or simulation of what-if scenarios, among many other alternatives. This tool will be used for something. Something that can be easy and fast, or something that can be complex, time-consuming, and long-term. Such usage will determine the rest of the life of this model. If the model is used for a one-time interpretation of parameters, not much life is ahead. But if the model is used for prediction and it’s intended to serve systems with online data collection, life is ahead of this newborn. What’s next then?
Maintaining the model
As we continue using the model, the conditions of the data that supported the training of the model will start changing. Right at that change the model starts experiencing changes as well. If we build a model for, say, churn prediction with a high prediction accuracy at the moment of training, in the near or far future the conditions or behavior of customers to be predicted start changing. This change is what challenges the prediction performance of our learned model. When those changes happen, our model enters a new phase which we call maintenance.
During the maintenance stage, we’ll likely need new data. With it, we can update the specification of the model. This is not different from taking another machine, say a car, and adjusting pieces when the machinery is not working properly. We won’t dive into strategies or solutions to perform model maintenance, but generally speaking, our models need to undergo a process of adjustment to bring them back to satisfactory performance.

Maintenance of machine learning models is not quite the same as retraining the models. Some models may be so simple that retraining them with updated data is just as simple. This may be the case with linear architectures or networks with very few layers and neurons. But when the models are so complex with large and deep architectures, the maintenance phase needs to be much simpler than the expense of retraining the model. This is one of the most important topics nowadays in the world of machine learning, as these tools can be so powerful, yet so expensive to maintain in the long run.
Once the model has been adjusted or updated, it is ready to go back to usage, so whatever process the model is serving can continue with the updated version. Our machine can continue to live. Nonetheless, this machine has already experienced changes. It has been used, consumed, if you will, and has been transformed into something slightly different from its original state. Like pencils, our models encounter those moments in which we need to sharpen their tips to preserve them so we can continue using them.
Transferring a model
Along the machine learning road, there’s an exit we might need to take: Transfer. I was once visiting stunning Iceland when I saw for the first time someone switching the tires of their car to go on the icy road. Then when they were back in the city, they switched back to the normal tires. When I started studying transfer learning, the concept became so clear while remembering the switch of tires for cars in Iceland. When new environments/domains come into play, our models enter a new phase known as transfer.

Just as the same car can be adapted to different grounds by switching the tires without having to purchase another separate car, we can add or adapt some parts of our models to serve new purposes in new domains without having to build new models. Transfer learning is another subfield of study in the machine learning literature aiming at optimizing the adjustment of models to simplify the effort of training models for new contexts. Popular examples of this are the models for image recognition. We train them with images of certain categories and later others transfer these models to recognize new categories of images. Many businesses use nowadays models such as RegNet, VGG, Inception, or AlexNet to adapt them to their own needs.
When we transfer a model, in a way, a new model is born with a lifetime of its own, separated from the original one. It will need maintenance just as the original model does. With that, we have gone from having one initial entity to probably creating a whole population of models. No doubt there’s indeed a lifetime behind these digital tools.

Does our model cease to live?
The short answer is: yes. They can indeed cease to live when, for example, their analytical performance is just systematically unsatisfactory or when they have become something so big and so different that the original models are just a thing of the past. As we said in the beginning, rocks, pencils, and also cars stop existing at some point. Models are not different from these things in that regard.
As real as it is that the model can become extinct, the answer to the question of when they reach this point is, to this day, the biggest question we want to answer within the machine learning research community. So many developments in monitoring the performance of machine learning and model maintenance are all related to the question of when the model is just not functional anymore.
One of the reasons why this answer is not a trivial one is because we constantly need labels to quantify how satisfactory the performance is. But the biggest paradox for machine and Statistical Learning is exactly that the labels are not available and we build these tools to predict them. Another reason is that it can be pretty subjective to define the limit of acceptance of change in performance. While scientists can propose some limits, businesses may have different tolerance levels.
Here are some points that the data scientist can also think of as part of answering this question (with the current open questions):
- Is the training data too outdated? (What is "too outdated")
- How similar is the current version to the original version of the model? (What is "similar"?)
- Has the variability of the input features and the relationship to the target variable completely drifted? (Covariate and concept drift, the two biggest topics of research in machine learning maintenance).
- Is the physical process where the model is deployed still in use? If the physical infrastructure does not support the deployment of the model anymore, this surely marks the end of its lifetime.
Ceasing to live for models is not necessarily a negative thing, it also resembles a path of their evolution. We need to make sense of its lifetime to keep our physical and digital systems up to date and with satisfactory performance.
TL;DR: Models are machines that emerge and transform over time
- When we train a model and start using it, it starts a lifetime road
- Simple uses of a model open simple lifetime roads. Complex uses make our models change.
- Changes come from the maintenance of our models. We need to update them or fix them. This gives us a new model
- Models can be transferred. New domains of use ask for an adjustment of the model.
- The life of the model is bound to its ongoing valid performance and its suitability in our systems.
Long live the model!