Lean and augmented data learning is the set of techniques that let a model learn well from limited data, by enlarging the training set with realistic variations or by reusing knowledge from a model trained elsewhere. The two main routes are data augmentation and transfer learning, and they are often combined.
This page is for teams whose problem has little historical or labelled data. It covers the methods, how they differ, how they relate to AI agents, and a short first-steps list.
Lean data learning aims to get useful models from small datasets. Augmented data learning adds variety to what you already have. Both respond to the same constraint: for many business problems, producing enough labelled examples is slow and costly. IBM's guide to transfer learning says it helps alleviate difficulties involved in acquiring large datasets, and notes that producing enough manually labelled data can be time-consuming and expensive.
The TensorFlow data augmentation tutorial defines it as a technique to increase the diversity of your training set by applying random but realistic transformations, such as image rotation. It lists layers such as random flip, rotation, contrast, crop and zoom.
Two cautions from the same page matter in practice. Augmentation should only be applied to the training set, and the tutorial does not claim that augmentation improves accuracy on its example. "Realistic" is a judgement you have to make for your own data.
The TensorFlow transfer learning tutorial explains that a pre-trained model is a saved network previously trained on a large dataset. You either use it as it is or customise it to a given task. The page names two routes: feature extraction, where the base model is frozen and only a new classifier is trained, and fine-tuning, where a few of the top layers of the base are also trained. It warns that fine-tuning is usually recommended when the training set is large and very similar to the original.
| Technique | What it does | Watch out for |
|---|---|---|
| Data augmentation | Adds random but realistic transformations to training examples | Apply only to training data; check the transformations make sense |
| Transfer learning: feature extraction | Reuses a frozen pre-trained model and trains a new classifier | Works best when tasks and data are similar |
| Transfer learning: fine-tuning | Also trains the top layers of the pre-trained model | Use a lower learning rate and train the new head first |
| Learned augmentation | A network learns which augmentations help most, called neural augmentation in this study | Results vary by dataset |
That study found that traditional transforms such as cropping, rotating and flipping are among the more effective strategies when training data is limited.
An agent is only as dependable as the models it calls. When a team has few examples of a rare event, a transferred and augmented model can be the practical starting point. ElixirClaw is XenonStack's AI agents computing platform, where such models can sit behind an agent. Record which data was real and which was augmented, so a reviewer can trace a decision back to its inputs.
To map one decision your agents make, and the data and policy it depends on, book a Context Workshop. See what happens in the workshop.
It is learning useful models from small datasets, often by reusing a pre-trained model through transfer learning.
It is a technique to increase the diversity of your training set by applying random but realistic transformations, such as image rotation.
No. The TensorFlow tutorial says data augmentation should only be applied to the training set.
Feature extraction freezes the pre-trained base and trains only a new classifier. Fine-tuning also trains the last layers of the base model.