Are there communities dealing with costs-vs-accuracy tradeoffs in Machine Learning?

Question

My group is working on a ML model that can work with little data (and bad accuracy) as long as there actually is little data available but can easily be extended as soon as said data becomes available (Think of interpolating existing models and then creating an individual predictor). This is due to a business requirement of the relevant application (new plants get installed over time but need to be integrated seamlessly). Transforming from a coarse yet cheap prediction to an accurate but expensive prediction takes labor effort that the company can invest as desired.

Are there research areas that take into account such "evolutionary transformation processes" which enable a tradeoff between costs and accuracy? What are the right keywords to look for?

I am looking for keywords/papers/communities.

Answer 1

Welcome to AI.SE ks.and1.

What you're describing could be related to several areas, but since you are looking only for some keywords, I'll suggest some:

1. You might be interested in anytime learning, which might better be called 'learning that can stop at any time'. A good example of this approach in machine learning is anytime learning for decision trees, as summarized in Esmeir & Markovitch's 2007 JMLR paper
2. Apart from this, it sounds like you are just describing the regular process of retraining a model when you get more data. For most algorithms, this just amounts to changing the input to your program so that it points at a larger file, containing more records than last time, and maybe buying a faster GPU or CPU to train the model with if it's taking too long with the extra data.

Hopefully, this can help you get started. If you have some more details about your problem, I might be able to suggest some more specific resources.