Bias is a systematic distortion in how data is selected, recorded, interpreted, or used to reach conclusions. Even accurate measurements can become biased when someone cherry-picks which observations to include. Subjective observations can add another layer because they reflect a person’s perception rather than an absolute fact. section 1.3 section 12.5
Bias appears in both data and model mechanics:
- Data bias occurs when a dataset overrepresents some groups, omits others, or encodes stereotypical roles. An image model trained mostly on examples of male doctors and female nurses may learn and amplify that pattern. section 12.5
- Perceptron bias is a learnable parameter that shifts the overall signal before activation. A master-gain control provides an intuition: it boosts or cuts the combined input independently of individual weights. section 9.2
How sampling creates bias
Suppose an AI system learns from photographs in which doctors are usually men and nurses are usually women. Those examples do not merely describe the dataset; they influence what the model predicts. A more diverse, representative dataset can reduce this distortion, although curating one is difficult. section 12.5 section 1.3
