AI Training (new Vocab term)

[riannella]

I would like to propose a new term for the Community Vocabulary.
Using (or not using) data for AI training is now a common use case in many sectors.

Definition: The computational process of teaching an Artificial Intelligence model to recognise patterns within Assets
Label: AI Training
Identifier: http://www.w3.org/ns/odrl/2/aitraining
Note: AI Training includes all techniques including Generative large language models
Included In: [use]
Class: [Action]

[joshcornejo]

Maybe we can reuse cawg'sterms?.

There are four pre-defined entries:
cawg.data_mining
Can any text or data content be extracted from the asset for purposes of determining “patterns, trends, and correlations.”
This would include images containing text, where the text could be extracted via OCR.
cawg.ai_inference
Can the asset be used as input to a trained AI/ML model for the purposes of inferring a result.
cawg.ai_generative_training
Can the asset be used as training data to an AI/ML model that could generate assets.
cawg.ai_training
Can the asset be used as data to train non-generative AI/ML models, such as those used for classification, object detection, etc.

[riannella]

Why not!

[besteves4]

I agree with the need to have a odrl:action for this. I would label it 'Train' though, to not make it specific only for AI. The specificity can come from then defining a purpose constraint for the training using, e.g., the terms mentioned in #147 (comment).

Then, maybe it would also be interesting to have a 'Test'/'Evaluate' action for policies related to testing/evaluating the models?

[riannella]

I would say that "Train" is more the "context of use" (like "education")... You don't actually "train" the asset (but that could be the purpose.)
However, the four CAWG terms are actual actions you perform/operate on assets.
(even though I don't like their labels or definitions !)

A good set may be:
AiGeneralTraining - Used to train general AI/ML models, such as those used for classification and object detection
Note: excludes GenerativeAI/LLMs
AiGenerativeTraining - Used to train an AI/ML model (eg LLMs) that could generate new Assets
AiAugmentation - Used as input to a trained AI/ML model to optimise the output result
AiMining - The process of determining patterns, trends, and correlations extracted from the Asset

Thoughts?

[joshcornejo]

I think the 4 extensions are sound, even if camel case looks weird :)

I also agree with "train" being too generic, contracts attempt to narrow down things, and it is easier to have a very large dataset that represents the taxonomy of actions than having to create extra artefacts to determine the context (this goes with my thinking around action refinements being unnecessary).

AWS access management has 158 actions that create concise context, eg:

• CreateAccessGrant
• CreateAccessGrantsInstance
• CreateAccessGrantsLocation