Training Methods

The side that decides, and the stage that reacts to it

Matching what an AI is capable of to what people actually want from it

Tracing error backward to split blame across the network

Losing old skill while learning something new

A saved snapshot of a model's state during or after training

Grouping similar things together with no labels at all

Learning by pulling alike things together and pushing others apart

When training stops improving and settles in place

Splitting data into pieces, checking each in turn, then averaging

Transforming data you already have to multiply how much of it you have

Info that shouldn't be in training slips in, inflating scores

A neural network standing in for a table of expected scores

Randomly resting part of a model during training

Stopping training once the held-back score stops improving

One full pass through all the training data

The balance between using what you know and trying something new

Skipping the data pool, gathering only what each side learned

Nudging a trained model further with new data

The ability to keep performing well on data never seen before

Mixing and nudging a batch of candidates to pick the best

The signal that says which way, and how far, to reduce error

Nudging values downhill toward less error, one small step at a time

A setting a person picks before training begins

Watching someone skilled and copying their moves exactly

Teaching a model to answer in the shape a request calls for

Pouring a big model's judgment into a smaller one

A line showing how the score changes as rounds or data grow

The multiplier that scales each training step

A spot that's stuck low, even though lower ground exists

Leaving the main body alone and training only a small add-on

A rule that measures how far off an answer is, as a single number

Finding rules from examples instead of writing them by hand

A penalty that squares how far off you were, then averages

Splitting data into chunks and processing one chunk per step

When the generator repeats a few winning outputs and variety fades

When AI keeps learning from its own output, variety quietly shrinks

The blur mixed in — and also where generation starts

The style of moving values once the gradient points the way

Memorizing the practice answers and failing on anything new

A standing rule for what to do in each situation

Learning the basics before anyone decides what the job will be

Writing down an expected score for every action and refining it

Naming a single number somewhere on a continuous scale

A device that fines a model whenever its values grow too large

Learning by trying things and adjusting to the score you get back

The single score that comes back after an action, good or bad

Polishing the texture of answers using what people prefer

The rule for how much better performance gets as scale grows

Learning by having the data make its own problems and answers

Learning from a little answered data and a lot that isn't

Learning a rule from examples that already carry the right answer

The process of nudging dials a little at a time from examples

Carrying a skill already learned over to a new task

Not getting even the practice data right

Finding structure in data with no answers attached

Data set aside for checking only, never used to train

When the fix-it signal almost fades out layer by layer, going back

A model of how the world works that you can run before acting