AI Exam Prep: 3D - Training

 

AI Practitioner Exam Prep - Training

Training (Learning) Types Terms

Continued Pre-training = FM gets latest raw U and updates its weights to teach domain knowledge. extends initial pre-training phase.

Deductive = using general rules to specific outcomes

Emergent = at large scales, these models develop skills that are not explicitly programmed into them.

Federated Learning = instead of bring data to central server (traditional), this brings model to the data. Good for data privacy and local compliance.

"Fine tuning" = improves already deployed pre-trained LM using small L (ex: industry-specific data) or input/output pairs often for behavior or task specialization. Most important task for fine tuning is labeling with accurate and relevant labels. Types are instruction tuning, RHLF, adapting models for specific domains, transfer learning, and continuous pretraining. 

"Generalization" = model's ability to apply knowledge from training on new unseen data. 

Inductive = using evidence to determine outcome. Builds a general model to predict future, unseen data.

Instruction Tuning = way to do fine tuning by putting ?/answer pairs.

"Masking" of Input = intentionally hiding parts of the input, forces models to understand context

Training in ML = iterative teaching a ML model to find patterns, make decisions, or generate content.

Transductive = predicts specific labels for fixed set of U by using both L training and distribution of the U test. Optimizes for performance of specific dataset.

Transfer Learning = takes existing pre-trained model on supervised task and then fine tunes.
Updating FM Weights = modifying numerical parameters that change how the model processes info (so allowing it to learn from new data)

Training (Learning) Parameters Terms (set on training environment itself)

Batch Size = Data processed per update. Small=faster iterations and generalization, large=stable and GPU efficiency.

Epochs = Neural networks. One epoch = Full pass through the dataset.

Hyperparameters =  Human-set dials.

"Learning Rate" = compares multiple trials to see improvement rate. controls the step size during optimization.

Parameters = Computer-learned weights.

Regularization parameter = increasing penalizes large weight during the training process, so makes smaller models so reduces chance of overfitting. training hyperparameter.