1. Bias in algorithms
Welcome! In this video, we will delve into how biases can be embedded in algorithms.
2. AI and data bias
Artificial intelligence, commonly referred to as AI, plays a pervasive role in people's lives, from deciding what content people see on their social media feeds to shaping the routes suggested by navigation apps for their daily commute. AI technologies are based on algorithms which are sets of rules designed to make predictions to support or even fully automate decision-making. A fundamental concern that arises is what happens if algorithms become biased against certain groups, such as gender or race.
3. Algorithmic bias
Data bias in the context of algorithms, often referred to as algorithmic bias arises when algorithms produce systematic and repeatable errors that result in unfair outcomes, favoring one group over another. It is an umbrella term that encompasses various types of biases that can emerge during the development and deployment of algorithms. Let's explore how algorithmic bias manifests.
4. Bias during algorithm training
Algorithmic bias is often initiated through bias in data collection, for example through the selection bias, and then it is reinforced by other bias types. If the data used to train an algorithm is not representative, the algorithm may learn and perpetuate those biases. For example, If the sample dataset used to train a facial recognition algorithm designed for security purposes does not adequately represent the diversity of the population it is meant to serve, the algorithm may not generalize well to all groups.
5. Bias in feature selection
Next, we have the feature selection bias. It refers to the introduction of bias in a machine learning model when certain features are chosen for inclusion based on criteria that may lead to unfair or discriminatory outcomes. This bias can manifest when selecting features that correlate with sensitive attributes, such as race, gender, or other protected characteristics. For example, an algorithm designed for predicting loan approvals may assume that income is the only relevant factor, ignoring other crucial aspects like educational background or credit history.
6. Evaluation bias
Using a non-representative dataset during the testing phase of an algorithm can lead to evaluation bias. This could give high overall accuracy scores, even if the algorithms were inaccurate for certain groups. Imagine we have an algorithm designed to recommend movies. Evaluation bias could occur if the algorithm was tested using a set of movies that only represented one genre, say action movies. If we then claim our algorithm performs great based on this limited test, we might be missing the fact that it doesn't work well for other genres like romance or comedy.
7. Automation bias
A concept closely linked to algorithmic bias is automation bias. Automation bias refers to the tendency of individuals to place undue confidence in the accuracy and reliability of automated systems without critically evaluating the information.
For example, consider a scenario where a medical diagnostic system is used to analyze medical images, such as MRIs, to detect abnormalities. Automation bias might occur if healthcare professionals consistently trust the system's results without independently reviewing or confirming the findings. In this case, if the automated system makes an error or encounters a scenario it wasn't trained for, the reliance on automation could lead to incorrect diagnoses.
8. Let's practice!
Great! Now that you've learned about algorithmic bias, let's put that knowledge into practice!