AI systems learn patterns from examples, and those examples are called training data. Training data can come from text, images, audio, videos, sensor readings, surveys, or human-made labels. The quality of the data matters because an AI model can only learn from what it is shown.
Careful data collection helps make AI tools more accurate, fair, and useful.
Key Facts
- Training data = examples used to teach an AI model to recognize patterns or make predictions.
- A dataset is usually split into training, validation, and test sets before a model is evaluated.
- Common split: 70% training, 15% validation, 15% testing.
- Accuracy = number of correct predictions / total number of predictions.
- Bias can appear when the data does not represent the people, situations, or examples the AI will face in real life.
- Privacy rules often require removing personal information such as names, addresses, faces, or ID numbers before data is used.
Vocabulary
- Training Data
- Training data is the collection of examples an AI system uses to learn patterns.
- Dataset
- A dataset is an organized group of data points, such as images, sentences, measurements, or labels.
- Label
- A label is the correct answer or category attached to a data example, such as marking an image as cat or dog.
- Bias
- Bias is a pattern in data or results that unfairly favors or excludes certain groups, examples, or outcomes.
- Validation Set
- A validation set is a portion of data used to tune a model while checking how well it performs on examples it did not train on.
Common Mistakes to Avoid
- Assuming more data is always better. A huge dataset can still produce poor results if it is noisy, mislabeled, outdated, or unfairly sampled.
- Mixing test data into training data. This is wrong because the model may appear more accurate than it really is by practicing on the same examples used for evaluation.
- Ignoring where the data came from. Data sources matter because copyrighted, private, low-quality, or unrepresentative data can create legal, ethical, and accuracy problems.
- Treating labels as automatically correct. Human labelers and automatic labeling tools can make mistakes, so labels must be checked for consistency and quality.
Practice Questions
- 1 A dataset has 10,000 images and is split into 70% training, 15% validation, and 15% testing. How many images are in each set?
- 2 An AI model makes 240 predictions on a test set and gets 204 correct. What is its accuracy as a decimal and as a percent?
- 3 A school trains an AI homework helper using only essays from advanced English classes. Explain one way this training data could create biased or unreliable results for the whole school.