Bias in AI: A primer

Bias in AI: A primer

While Artificial Intelligence (AI) systems can be highly accurate, they are imperfect. As such, they may make incorrect decisions or predictions. Several challenges need to be solved for the development and adoption of technology.
One major challenge is the bias in AI systems. Bias in AI refers to the systematic differences between a model’s predicted and true output. These deviations can lead to incorrect or unfair outcomes, which can seriously affect critical fields like healthcare, finance, and criminal justice.

AI algorithms are only as good as the data they are trained on. If the data is biased, the AI system will also be. This can lead to unfair and unequal treatment of particular groups of people based on race, gender, age or sexuality. Adding to these problematic ethical implications is the potential for misuse of the technology by corporations or politicians.
Several types of bias can occur in AI.

These include:

• Selection/ Sampling bias: This occurs when the total dataset correctly represents the population, but sample of data used to build the model is only partially representative of the entire population. For example, if a model is trained on data from a single demographic group, it may not accurately predict outcomes for other groups.
• Algorithmic bias: This occurs when the algorithms used to build the model are inherently biased. For example, a model that uses gender as a predictor variable may be biased against a particular gender. As a result, the model learns the incorrect representation of data and its relationships within and with the target variable.
• Human bias: This occurs when the data used to train the model is biased due to human error or bias. For example, if a dataset contains mostly positive examples, a model trained on this data may be biased toward predicting positive outcomes.
• Representation bias: This occurs when certain groups or individuals are underrepresented in the overall dataset from which the training data is taken. For example, a model trained on data from a predominantly white population may not accurately predict outcomes for people of color. Here the data is not inclusive of the anomalies, outliers and diversity of the population. This is distinct from sampling bias because it concerns the entire underlying dataset–not just the sample.

Recognizing and addressing bias is important, as it can lead to unfair and potentially harmful outcomes. Some steps that can be taken to mitigate bias in general include:

• Ensuring that the training data is representative of the entire population
• Using algorithms that are less prone to bias
• Regularly checking for discrimination in the model’s predictions
• Ensuring that the model is fair and unbiased in its treatment of different groups

By taking these steps, we can work towards building machine learning models that are more accurate, fair, and unbiased.

Additionally, bias can occur in a model at various stages of the model-building process.

These include:

• The data gathering stage: Bias could be present in the data itself due to underlying disparities between groups or due to the data gathering process. For instance, data collected through a survey targeting a specific demographic is non-representative of the global population.
• The labeling stage: Bias could be introduced at the labeling stage if the labeling process is biased against certain groups. Human annotators can have individual biases at different levels. This can skew the data that will be picked up during model training.
• The modeling stage: Bias can also be introduced at the modeling stage as different models have learned information in different ways. For instance, some models may be better than others at learning examples that may occur more infrequently for the minority group.

Let’s get into more detail and learn how to mitigate each one of the types of bias in AI.

Mitigating selection/sampling bias

Selection/sample bias refers to the systematic differences between the training data used to build the model and the true population. This can occur when the training data sample is representative of only some of the population, resulting in an inaccurate or generalizable model.

With this bias type, the data has the complete representation of the entire population/scenario but ends up underrepresenting one or more demographics or samples during selection. There are several ways to mitigate sample bias in AI. These include:

• Stratified sampling: This sampling method ensures that the training data is representative of the entire population by dividing the population into smaller groups (strata) and selecting a representative sample from each group.
• Oversampling or undersampling: If certain groups or classes are underrepresented in the training data, oversampling can be used to increase the number of examples from these groups. Undersampling can be used to decrease the number of samples from groups that are overrepresented.
• Data augmentation: Data augmentation involves generating synthetic examples from the training data to increase the size of the dataset. This can help the model learn more about the relationships between different features and improve its generalizability.

Mitigating algorithmic bias

Algorithmic bias refers to the inherent biases present in the algorithms used to build the model. These biases can arise from various sources, including the assumptions built into the algorithm and the data used to train the model.

There are several ways to mitigate algorithmic bias in machine learning. These include:

• Using algorithms that are less prone to bias: Some algorithms, such as decision trees and logistic regression, are more prone to bias than others, such as support vector machines and neural networks. Choosing an algorithm that is less prone to bias can help mitigate the risk of bias in the model.
• Using a diverse and representative dataset: Using a diverse and representative dataset to train even a biased algorithm can also help reduce algorithmic bias by ensuring that the model is exposed to a wide range of sufficient examples.
• Fairness metrics: Various fairness metrics can be used to evaluate the fairness of a machine learning model. These metrics can help identify bias in the model and suggest ways to mitigate it.
• Leveraging debiasing techniques: There are a few techniques that can be used to reduce bias in machine learning models, such as preprocessing the data to remove sensitive variables, applying regularization to the model, and using counterfactual data.

By taking these steps, it is possible to reduce algorithmic bias and build fairer and more accurate machine learning models.

Mitigating human bias

Human bias refers to the prejudices in the data used to train the model due to human error or opinion. These biases can arise from various sources that the human interacted with the model design directly or how the data was collected, labeled, or even processed.

There are several ways to mitigate human bias in AI. These include:

• Automated data labeling techniques: Automated data labeling can help minimize human bias by removing the need for human annotators. This can be done using techniques such as active learning, which allows the model to select examples for labeling based on its current performance.
• Using debiasing techniques: Various human-intensive techniques can reduce bias in machine learning models. These include preprocessing the data to remove sensitive variables, applying regularization to the model, and using counterfactual data.
• Monitoring and assessing the model’s performance: It is essential to monitor and evaluate the model’s performance to ensure that it is not exhibiting bias. This can be done using various fairness metrics and comparing the model’s predictions to the true outcomes.

Mitigating representation bias

Representation bias refers to the biases present in the training data due to the underrepresentation of certain groups or individuals. This can occur when certain groups or individuals are not included in the training data or are significantly underrepresented.

With this bias type, the data itself is incompletely representing the entire population/scenario. Thus the model is trained on underrepresented scenarios leading to bias. There are several ways to mitigate representation bias in machine learning. These include:

• Using a diverse and representative dataset: Ensuring that the training data is representative of the entire population can help reduce representation bias by ensuring that the model is exposed to various examples from different groups.
• Use oversampling or undersampling: If certain groups or classes are underrepresented in the training data, oversampling can be used to increase the number of examples from these groups. Undersampling can be used to decrease the number of examples from groups that are overrepresented.
• Use data augmentation: Data augmentation involves generating synthetic examples from the training data to increase the size of the dataset. This can help the model learn more about the relationships between different features and improve its generalizability.
• Use fairness metrics: Various fairness metrics can be used to evaluate the fairness of a machine learning model. These metrics can help identify bias in the model and suggest ways to mitigate it.

In this blog, we focused on bias in AI. We decoded some of the crucial questions around this topic, from types of bias to preventive methods. We hope you use the tips discussed to design fair and responsible AI systems.