Underfitting

As explained in Generative AI in Action

Underfitting is a modeling condition in which a model fails to learn enough structure from its training data to perform well on new examples.

Underfitting occurs when a fine-tuned model fails to learn the underlying patterns in its training data well enough, leaving both training and validation loss high. It indicates insufficient learning rather than poor generalization caused by memorizing the training set. section 9.3.4

Use training metrics to distinguish underfitting from overfitting and guide fine-tuning decisions:

• Check whether both training and validation loss remain high; this pattern signals that the model is not learning enough structure. section 9.3.4
• Consider additional training when loss remains high or validation loss has not stabilized, while accounting for training-step count and loss scale. section 9.3.4
• Adjust dataset quantity and quality based on monitored performance; task complexity, data quality, computational resources, and time affect the balance. section 9.1.2
• Separate underfitting from overfitting: rising validation loss alongside falling training loss suggests overfitting, not insufficient learning. section 9.3.4

Reading a fine-tuning curve

Suppose a GPT-3.5 Turbo fine-tuning run shows persistently high training and validation loss. Treat the result as possible underfitting, then review training steps, loss scale, and dataset adequacy before judging the model. section 9.3.4 section 9.1.2

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

Inspect the training-loss figure to compare learning progress with the expected decrease and stabilization of loss. section 9.3.4

How Generative AI in Action covers it

How deep it goes

Interprets training and validation loss patterns, discusses training duration and data adequacy, and uses no mathematical derivation.

Examples

Uses GPT-3.5 Turbo training-loss curves and monitored dataset adjustments as concrete diagnostic examples.

What it assumes

Assumes familiarity with training and validation datasets and their loss metrics when interpreting fine-tuning results.

What you can do after

Diagnose underfitting with training and validation loss, then evaluate whether more training or dataset changes are warranted.

How it connects

Connects underfitting with fine-tuning, training loss, validation loss, overfitting, generalization, and model evaluation.

liveBookmanning.com

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

Also explained in

• 

  liveBookmanning.com
  Grokking Machine LearningA model whose complexity is insufficient to capture the relevant structure of its dataset.
• 

  liveBookmanning.com
  Machine Learning System DesignA condition in which a model is too simple to capture important relationships between features and the target variable, resulting in poor performance on both training and new data.
• 

  liveBookmanning.com
  Grokking AI Algorithms, Second EditionUnderfitting is a condition in which a model is too simple or insufficiently trained to perform well even on its training data.
• 

  liveBookmanning.com
  Deep Learning with R, Third EditionUnderfitting is a modeling condition in which a model is too simple or insufficiently trained to capture patterns needed for good performance.
• 

  liveBookmanning.com
  Machine Learning for Tabular Data: XGBoost, Deep Learning, and AIUnderfitting is a modeling failure in which a model is too simple to capture important patterns in the data.
• 

  liveBookmanning.com
  The AI Pocket BookA failure mode where a model is too simple to capture the training pattern and performs poorly.

Compare how 7 books explain underfitting →

Related concepts

• overfittingcontrasts with · Overfitting is a modeling condition in which a model fits patterns or idiosyncrasies in its training data too closely and consequently generalizes poorly to new data.
• genetic algorithmaddressed by · A genetic algorithm is an evolutionary optimization method that improves a population of candidate solutions through fitness-based selection and variation, often including recombination and mutation.
• bias-variance tradeoffcontrasts with · The bias-variance tradeoff is the balance between systematic error from restrictive model assumptions and error from sensitivity to variation in sampled data.
• learning curvemeasured by · A learning curve is a representation of how performance or proficiency changes as experience, training progress, or training data increases.
• expressive poweraddresses · Expressive power is the capacity of a model, language, or system to represent or convey a range and complexity of functions, patterns, distinctions, or ideas.

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.