According to [1], in MLOps, continuous training is

a new property, unique to ML systems, that's concerned with automatically retraining and serving the models.

While lifelong/incremental learning mainly studies how to incrementally learn rather than retrain. [2]

Lifelong Machine Learning or Lifelong Learning (LL) is an advanced machine learning (ML) paradigm that learns continuously, accumulates the knowledge learned in the past, and uses/adapts it to help future learning and problem-solving.

I can see some links or conflicts between the two but cannot explicitly explain, and I asked an author in the second link about this issue and he said that the two are complementary. I wonder how will the two help each other? Or will this kill that?


2 Answers 2


Lifelong learning and MLOps are indeed complementary.

Lifelong learning (LL) can be defined as the set of learning algorithms and models that can deal with more and more data and/or tasks without forgetting (completely) the previous one and (usually) without fully retraining the model with all data that you have available now. So, in LL, we attempt to mimic the way humans continually learn (different tasks) throughout their lives and transfer knowledge/skills acquired in one task to other tasks (e.g. from the task of walking to the task of running) [1].

MLOps is the application of DevOps to the machine learning or data science context. In other words, it is more a software engineering practice, where you define a pipeline (or sequence) of tasks that need to be done (semi-)automatically until and after you deploy your software. So, MLOps is just DevOps, but, in addition to the common practices in DevOps, like continuous integration and continuous deployment, you also have ways to deal with the ML part of the software, like automatic retraining as more data is available, and automatic evaluation of the (new) trained model on a test dataset.

They are complementary, because MLOps is concerned with the automatization of the data science process (i.e. collection and cleansing of data, training of the model with the new data, evaluation, deployment of the new trained and evaluated model to a production server), while LL is (only/usually) concerned with the development of learning algorithms and models that can cope with new tasks and data (without being completely reconfigured or retrained), so it has nothing or little to do with continuous integration and continuous deployment. In MLOps, given that you may (automatically) retrain the model as more data is available, you may think that this is a form of lifelong learning, but you completely retrain the model and you may not be able to cope with different tasks, but you could indeed use LL techniques in MLOps.


To combat data distribution shifts, especially sudden shifts, or to adapt to rare events, a deployed model should be retrained continually.

Incremental learning is one of the two manners of how a model is retrained(continual learning): 1) stateless retraining; and 2) stateful retraining(fine-tuning or incremental learning), both of which are largely an infrastructural problem, that is, MLOps.

Most companies do stateless retraining—the model is trained from scratch each time.

As of stateful retraining:

As of today, stateful training is mostly applied for data iteration, as changing your model architecture or adding a new feature still requires training the resulting model from scratch. There has been research showing that it might be possible to bypass training from scratch for model iteration by using techniques such as knowledge transfer (Google, 2015) and model surgery (OpenAI, 2019). According to OpenAI, “Surgery transfers trained weights from one network to another after a selection process to determine which sections of the model are unchanged and which must be re- initialized.” Several large research labs have experimented with this; however, I’m not aware of any clear results in the industry.

Since modeling is only one part of the MLOps pipeline: ![enter image description here

Incremental learning also heavily relied on MLOps, such as for data labeling, evaluation, etc.


  1. Designing Machine Learning Systems

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .