liveBook · Manning

Bias

As explained in Build AI Drug Discovery Pipelines

Statistical bias is systematic error caused by a model's simplifying assumptions that prevent it from capturing the true structure of the data.

Statistical bias is systematic error from simplifying assumptions that prevent a model from capturing the data’s true structure. A high-bias model is too simple, underfits, and usually has relatively high training error. The goal is sufficient complexity to capture real patterns without fitting noise. section 4.4 section 4.4.1

Use bias alongside variance to interpret model behavior:

  • Compare training and test errors: high bias commonly produces high training error, while high variance produces a large gap between training and unseen-data performance. section 4.4.1
  • Inspect learning curves: if errors converge at a high level, collecting more data is unlikely to solve the problem; a more complex model may help. section 4.4.2
  • Inspect validation curves: low-complexity hyperparameters create high bias, while excessive complexity creates high variance; choose the region with minimum test error. section 4.4.3
  • Control the trade-off with regularization, which can lower variance while increasing bias; decision-tree limits and pruning provide concrete examples. section 4.4.2 section 5.3.1

Polynomial models for solubility prediction

A second-order polynomial may have high bias because it cannot represent a more complex target function, so it underfits. A twentieth-order polynomial has lower bias but can fit noise and generalize poorly because its variance is high. The simpler model can nevertheless achieve lower test error, showing why complexity alone does not determine performance. section 4.4.1

Figure 4.18 How high bias or high variance impact a model’s sensitivity to the training data it learns from.

Use the figure to compare how high-bias and high-variance fits respond to changes in the training data. section 4.4.1

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines bias, relates it to training and test error, and diagnoses it with learning and validation curves.
Examples
Uses second- and twentieth-order polynomials, RANSAC and SVR learning curves, and regularized decision trees.
What it assumes
Assumes basic model evaluation and introduces bias through polynomial fitting, training error, test error, and model complexity.
What you can do after
Diagnose underfitting with error curves and adjust model complexity, data collection, or regularization.
How it connects
Connects bias to variance, underfitting, overfitting, regularization, hyperparameters, learning curves, and validation curves.

Also explained in

Compare how 38 books explain bias →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build AI Drug Discovery Pipelines. The section links point to the parts of the book it draws on.