Catastrophic forgetting in online models
Although catastrophic forgetting was initially identified as a problem for neural networks, you can imagine that online machine learning has the same problem of continuous re-learning. The problem of catastrophic forgetting, or catastrophic inference, is therefore also present and needs to be mastered.
If models are updated at every new data point, it is expected that coefficients will change over time. Yet as modern-day machine learning algorithms are very complex and have huge numbers of coefficients or trees, it is a fairly difficult task to keep a close eye on them.
In an ideal world, the most beneficial goal would probably be to try and avoid any wrong learning in your machine learning at all. One way to do this is to keep a close eye on model performance and keep tight versioning systems in place to make sure that even if your model is wrongly learning anything, it does not get deployed in a production system. We will go into this...