training

Direction test: training loop is the whole process, backpropagation is a structural sub-step (the backward pass). The training loop contains backpropagation as one of its constituent parts — like a book contains chapters (Law 10). If you remove backpropagation from a training loop, the loop stops operating as a learning system (forward pass only). Whole-part relationship, not dependency.

A training loop is a specific kind of procedure: a structured sequence of steps repeated iteratively to adjust model parameters during machine learning training. A competent speaker would call a training loop 'a procedure.'

Safety-training is a specific kind of training focused on safety knowledge, procedures, and competencies. It is a domain-specific instance of the general training construct.

coaching INSTANCE_OF training per Law 9: 'coaching is a specific kind of training' focused on individual development, performance improvement, and goal-setting rather than skill acquisition. Nearest kind is training. A competent speaker would call coaching 'a form of training.'

Fine-tuning is a specific kind of training — the specific case of further-training a pre-trained model on a task-specific dataset. A competent ML practitioner calls fine-tuning 'a kind of training.' Tests specific→general INSTANCE_OF.

Knowledge distillation is a specific kind of training method in machine learning. A competent speaker would call it 'a training technique.' Nearest kind: training, not model compression (which is a goal, not a training method).

Batch normalization is built and maintained for the sake of training — it stabilizes training dynamics and accelerates convergence. Its designed purpose is to further training's operation (Law 8d: for whose sake? training is the master).

Learning rate scheduling operates during training — it adjusts the learning rate as part of the training process. The removal test: remove training, learning rate scheduling stops operating entirely (nothing to schedule for). Files present-tense dependency per Law 8.

Learning rate is a parameter that only operates within the context of training. Remove training — the concept of learning rate ceases to function operationally. The learning rate specifies step size for parameter updates during training; without training there is nothing to parameterize. Constitutive dependency: learning rate is a dimension of training, not an independent concept.

DERIVED_FROM: training as a concept existed first; training loops are a specific implementation that came later. Historical chronology test passes — the general concept of training predates the specific construct of a training loop.

An optimizer is built and maintained for the sake of training — its designed purpose is to adjust model parameters during training. The servant (optimizer) points at the master (training), per Law 8d.

Early stopping halts training based on monitored metrics — it is built and maintained for the sake of training quality and resource efficiency. The designed purpose is to improve training outcomes, per Law 8d.

curriculum learning is built and maintained for the sake of training — its designed purpose is to improve the training process by ordering examples for optimal learning. SERVES direction: servant (curriculum learning) → master (training). For whose sake? Training.

Backpropagation is an algorithm specifically designed and maintained for the sake of training neural networks. Its sole purpose is to compute gradients efficiently to enable weight updates during training. Servant→master: backpropagation serves training.

Working-memory-training is a specific kind of training: it is a structured program designed to improve working memory capacity through repeated exercises. Direction tested (specific→general): working-memory-training IS A training method.

Chronological and conceptual priority (Law 7): the concept of training models predates and fed into the identification of underfitting as a distinct failure mode. Underfitting only makes sense as a concept within the broader practice of training.

The removal test: in-context learning operates on a model's fixed weights after training. Remove training — the model has no learned representations and cannot perform in-context learning. The training establishes the representational capacity that makes the mechanism work.

Curriculum learning is a training strategy that organizes the order of training examples to improve learning efficiency. Remove training and curriculum learning stops operating — it is a method OF training. The removal test passes.