Ch 3: Data and Fairness in AI Class 8 Notes: Important Points
Ch 3: Data and Fairness in AI Class 8 Notes: Important Points
MIND MAP (Click to download)
Introduction
AI systems learn from the data provided during training. If this data is balanced, diverse, and accurate, the AI performs well. However, if the training data contains bias or is unbalanced, the AI may make unfair or incorrect decisions.
In this chapter, we will understand the importance of data in AI, how AI learns from patterns, why fairness matters, and how biased data can affect AI predictions.
Data in AI
One of the artists’ groups in Seoul, Korea, collected photographs of clouds that looked slightly similar to human faces. They gave it to an AI face-detection program and very surprisingly, the program reported all the images as face detected!!
The above example let us realize that:
- AI looks for patterns they have learned from the training dataset.
- The training data influences the AI decisions.
- It becomes crucial to analyse the balance in the training data.
Role of Data
Let us consider that we have to train an AI model to identify a cricket bat and badminton racket. The AI model will be provided with many images of a bat and a racket. AI carefully studies these images and looks for patterns such as the long, flat shape of a cricket bat, the oval frame and strings of a badminton racket. Over time, AI learns to distinguish between these two objects based on these patterns.
Most training images are of cricket bats compared to a badminton racket. Now, if we show the new image, what will the AI guess?
It will guess a cricket bat for most of the time. This happens because the AI learns only from the data
it is given.
NOTE: When one type of object appears more often in the training data, the AI becomes better at recognizing that object. It becomes less accurate for objects with fewer examples. The correct image recognition will depend heavily on the balance in the training data.
Importance of Data
AI does not think like humans. It only learns from the data we give for training. If the data is good and balanced, the AI works well. However, problems arise when certain groups of data are overrepresented or underrepresented in the training.
Automatic translation in AI sometimes produces gender stereotypes. for example
But the Hindi sentence does not specify gender. Training data often contained more examples of male doctors. AI will learn stereotypes present in the training data.
Understanding Bias
Computer scientist Joy Buolamwini tested AI-technology to recognise face. She discovered that some systems worked very well for light-skinned men but made more mistakes for women and people with darker skin tones
Reason: the systems were trained using datasets that contained more photos of certain groups than others.
This showed that AI systems can behave unfairly when the training data does not represent everyone.
Bias in Dataset
1. Some data does not represent everyone equally. This is called bias.
2. Bias means the data is unbalanced or unfair.
AI systems are used (or will be used) in many places
Now a days AI system can be used for recommending videos, approving loans, helping doctors, selecting job applicants. But if the data contains bias, the AI may make unfair decisions. for example
A school is building an AI program that recommends sports to students. The training data is given below:
| Sports | Boys | Girls |
| Cricket | 55 | 15 |
| Badminton | 10 | 20 |
From the above training data that the AI program may notice that most boys played cricket, and most girls played badminton.
NOTE: The above example shows how AI can learn human biases from historical data if developers are not careful. AI should be built to help everyone, not just a few people.
How can we make AI Fair?
- Bias check: Test the AI-programs to see if they recognise/treat various categories differently.
- Human supervision: Human review the AI decisions instead of trusting them blindly.
- Transparency: Understanding how an AI system works and how it makes decisions.
Activity: Supervised Learning & Bias?
- Go to https://teachablemachine.withgoogle.com and click on ‘Train models on images’
- Click ‘Add samples’, then ‘file’ & ‘Choose files’
- Locate the Teachable Machines Cat-Dog Dataset in the Drive.
- Click ‘Dogs’, then ‘training dataset’. Click ‘open’ to add an image to the training set. Be sure to
add all images at once. - After all of the images have been added, click Add samples. Repeat with cat training images
One Page Revision(click to download)
Disclaimer : I tried to give you the Correct Notes of “Ch 3: Data and Fairness in AI Class 8 Notes“ , but if you feel that there is/are mistakes in the Notes of “Ch 3: Data and Fairness in AI Class 8 Notes“ given above, you can directly contact me at csiplearninghub@gmail.com. NCERT Book and Study material available on CBSE official website are used as a reference to create above “Ch 3: Data and Fairness in AI Class 8 Notes“ All the screenshots used in above article are taken from NCERT Book and Study material available on CBSE official website.