bact
commented
3 weeks ago In the absence of an actual measurement, an estimation is acceptable -- at least according to the Section 1 of Annex XI of EU AI Act.
The record is from the model provider side, based on their configuration (if it is about training, it can be documented in informationAboutTraining property).
The user/deployer of the model can have different numbers, they can record theirs in a new object/new SPDX document. For example, after they finetuned a pretained model, they will have another set of energy consumption records.
Description of 'finetuningEnergyConsumption' says that it's the energy consumed for finetuning an AI model. Description of 'energyConsumption' says that it's the energy consumed for training an AI model. Description of 'inferenceEnergyConsumption' says that it's the energy consumed during inference.
An AI model could go through multiple fine tunings in its lifecycle. The amount of energy consumed would vary from instance to instance, and won't be constant to be recorded. Training of an AI model can happen multiple times, so the energy consumed will not be the same for every instance. Likewise, energy consumed during inference would vary depending on the amount and nature of the data that's being processed.
These properties attempt to capture these value (as a constant). I am not sure if this is correct. If it is indeed correct, then more explanation is needed as to what these property mean and is meant for.