Data Augmentation for AI Training

Training an AI model usually depends on data. Sometimes the data you need is scarce, expensive to gather, or missing the cases you care most about. A common response is to create data artificially rather than waiting to collect more of it.

The idea of creating data artificially is older than AI. Data synthesis has a long history in software testing, gaming, and robotics, and the techniques developed in those fields are the foundation for the approaches covered in this article. Before asking how the process is done for AI, it helps to understand the older methods that it builds on.

There are two broad ways to get a set of data items. In manual generation, a person creates each item by hand. Using algorithms to generate data is called procedural generation. In procedural generation, a program produces the items according to a defined procedure, so the same procedure can produce many items without a person writing each one.

The clearest everyday illustration comes from gaming. Procedural generation is commonly used in games to create levels, maps, items, and characters on the fly. The content is produced as the game runs rather than being stored in advance. The source notes that most data generation techniques used in these industries can be applied to AI. That is the bridge: a method that builds game worlds can, in principle, help build training data.

Rule-based data synthesis is a traditional approach that produces data from predefined rules. The rules are written before generation begins. The generator does not learn them from examples; it applies them. Each output is therefore a product of the rules, and understanding a rule-based generator means understanding its rules.

Because the rules are explicit, a rule-based approach makes the structure of the output easy to control. If a rule says a value must fall in a certain range, every generated value falls in that range. The same explicitness is also the limit: the generator can only produce what the rules describe. A rule-based generator's output is bounded by its rules, so its structure is predictable, but its coverage is limited to what those rules describe.

The following example uses rules invented for teaching purposes; the source does not describe these specific rules. Suppose a generator has two rules. Rule one: the next value is the previous value plus 3. Rule two: a value is kept only if it is at most 10. The generator starts at 2 and produces values until a value fails rule two.

Simulation-based data synthesis is the second traditional approach. Instead of writing data-producing rules directly, it generates data by running a simulation. The source describes this approach briefly as a traditional technique alongside rule-based synthesis, and it places robotics among the fields where data synthesis has a long history.

A simple way to picture the difference is this. A rule-based generator asks, for each item, what the rules allow. A simulation-based generator lets a modeled process unfold and records what it produces. The data comes out of the run of the model rather than being assembled field by field.

Traditionally, data synthesis and augmentation have followed two approaches: rule-based and simulation-based. A newer method, made possible by advanced AI models, uses AI itself to synthesize data. The diagram below places these three under the general idea of data synthesis. Read it as a family tree of methods rather than a sequence of steps.

The first two approaches rely on rules or simulations that a person designs. The third relies on an AI model to produce the data. At the conceptual level, that is the main thing to hold onto before the AI-powered method is examined in detail.