Artificial intelligence learns from training data by studying many examples and looking for patterns that connect inputs to correct outputs. For example, an image model may see thousands of labeled pictures of cats, dogs, cars, and bikes. The model does not memorize every picture perfectly, but it adjusts itself to make better guesses on new examples.
This matters because the quality and variety of the training data strongly affect how useful and fair an AI system becomes.
Inside many AI systems, a neural network uses adjustable numbers called weights to transform input data into predictions. During training, the model makes a prediction, compares it with the correct answer, measures the error, and updates its weights. This process repeats for many rounds so the model gradually improves.
A well trained model can generalize, which means it can make reasonable predictions on examples it has not seen before.
Key Facts
- Training data = examples plus the correct labels or answers the model should learn from.
- Prediction error can be measured with a loss function, such as loss = predicted value - true value for a simple error.
- Weights are adjustable numbers inside a model that control how strongly one feature affects a prediction.
- A training loop often follows: input data, predict, calculate loss, adjust weights, repeat.
- Learning rate controls update size, often written as new weight = old weight - learning rate x gradient.
- Good models should perform well on test data, not just on the training data.
Vocabulary
- Training data
- Training data is the collection of examples an AI model studies to learn patterns.
- Label
- A label is the correct answer attached to a training example, such as cat, spam, or not spam.
- Neural network
- A neural network is an AI model made of connected layers that process information using adjustable weights.
- Weight
- A weight is a number inside a model that changes during training to make predictions more accurate.
- Loss function
- A loss function is a rule that measures how far a model's prediction is from the correct answer.
Common Mistakes to Avoid
- Thinking AI understands like a human, which is wrong because most AI systems learn statistical patterns from data rather than real-world meaning or personal experience.
- Using only easy or similar training examples, which is wrong because the model may fail when it sees new examples that are different from the training set.
- Judging a model only by training accuracy, which is wrong because a model can memorize training data and still perform poorly on new data.
- Setting the learning rate too high, which is wrong because weight updates can jump past the best solution and make training unstable.
Practice Questions
- 1 A model classifies 200 test images and gets 174 correct. What is its accuracy as a percent?
- 2 A model starts with a weight of 0.80. The learning rate is 0.10 and the gradient is 0.30. Using new weight = old weight - learning rate x gradient, what is the new weight?
- 3 A school trains an AI to recognize backpacks using only photos of red backpacks. Explain why the AI may perform poorly on blue or black backpacks and how the training data could be improved.