Generalization

As explained in Generative AI in Action

Generalization is a model's ability to perform effectively on examples that were not present in its training data.

Generalization is a model’s ability to perform effectively on examples that were absent from training. In fine-tuning, it matters because a model can learn the training set without learning patterns that transfer to new inputs. section 9.3.4 section 9.1.2

Use training and validation loss to assess whether fine-tuning is learning transferable patterns:

• A good fit shows both losses decreasing toward stability, with only a small gap between them. section 9.3.4
• Overfitting is likely when training loss decreases while validation loss increases; the model is fitting training data but generalizing poorly. section 9.3.4
• Underfitting is likely when both losses remain high, indicating that the model has not learned the underlying patterns sufficiently. section 9.3.4
• Judge loss by its scale, training-step count, and stability; batch-level fluctuations are normal when the overall loss improves. section 9.3.4

Reading a fine-tuning run

A fine-tuning run may show steadily declining training loss. If validation loss also declines and stabilizes nearby, the model is likely generalizing well. If validation loss begins rising, investigate overfitting rather than simply adding training steps. Figure 9.3 provides a training-loss view for GPT-3.5 Turbo fine-tuning. section 9.3.4

Figure 9.3 Training loss when fine-tuning GPT-3.5 Turbo

Inspect the training-loss figure to compare progress across training steps and relate loss behavior to generalization. section 9.3.4

How Generative AI in Action covers it

How deep it goes

Explains training-validation loss patterns, overfitting, underfitting, batch noise, and stability without deriving model internals.

Examples

Uses training-loss curves for GPT-3.5 Turbo fine-tuning and contrasts improving, overfit, and underfit loss patterns.

What it assumes

Assumes familiarity with training and validation datasets while explaining their relationship through loss behavior.

What you can do after

Diagnose fine-tuning quality by comparing training and validation loss and recognizing overfitting or underfitting.

How it connects

Links fine-tuning, training loss, validation loss, overfitting, underfitting, and model evaluation.

liveBookmanning.com

Read Generative AI in Action on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Deep Learning with Python, Third EditionThe ability of a model or agent to perform well on novel situations different from those seen during training or prior exposure.
• 

  liveBookmanning.com
  Deep Learning with R, Third EditionGeneralization is a model's ability to perform accurately on previously unseen inputs from the same underlying domain as its training data.
• 

  liveBookmanning.com
  Natural Language Processing in Action, Second EditionGeneralization is a model's ability to produce useful predictions for previously unseen inputs, including inputs outside its training examples.
• 

  liveBookmanning.com
  Sutskever's ListGeneralization is a model's ability to maintain predictive performance on inputs beyond those used for training.
• 

  liveBookmanning.com
  Deep Learning with PyTorch, Second EditionThe ability of a trained model to perform well on data not seen during training by capturing underlying patterns rather than memorizing specifics.
• 

  liveBookmanning.com
  Evolutionary Deep Learning: Genetic algorithms and neural networksGeneralization is a model's ability to perform well across environments or data conditions beyond those directly optimized during development.
• 

  liveBookmanning.com
  The AI Pocket BookThe ability of a model to perform well on new inputs not seen during training.
• 

  liveBookmanning.com
  Machine Learning Algorithms in DepthThe ability of a learned model to perform accurately on previously unseen examples.
• 

  liveBookmanning.com
  Effective Data AnalysisThe extent to which findings from a study apply to populations, settings, or conditions beyond those directly examined.
• 

  liveBookmanning.com
  Knowledge Graphs and LLMs in ActionGeneralization is the process of deriving a broader model or conclusion that applies beyond the observed examples.
• 

  liveBookmanning.com
  Grokking AI Algorithms, Second EditionGeneralization is a model's ability to perform effectively on new data drawn from the same underlying task or distribution.
• 

  liveBookmanning.com
  Grokking StatisticsGeneralization is the ability of a fitted model to perform appropriately on new data beyond the observations used to fit it.

Compare how 13 books explain generalization →

Related concepts

• validation lossmeasured by · The value of a loss function computed on held-out validation data to assess a model’s predictions.
• Abstraction & Reasoning Corpus for Artificial General Intelligencemeasured by · The Abstraction and Reasoning Corpus for Artificial General Intelligence is a benchmark of visual reasoning tasks designed to assess abstract rule induction and generalization from examples.
• metricmeasured by · A metric is a quantitative measure used to characterize or assess a property, condition, behavior, performance, or outcome.
• optimizationcontrasts with · Optimization is the systematic selection or adjustment of values, actions, or system designs to improve a specified objective, often subject to constraints.
• memorizationcontrasts with · Memorization is the retention of specific information or examples in a form that enables their later recall or reproduction.
• parameter countcontrasts with · Parameter count is the number of numerical parameters in a model that are learned or adjusted during training.

This explainer was generated by Manning's knowledge graph from the text of Generative AI in Action. The section links point to the parts of the book it draws on.