Learning from Examples

Machine learning (ML) is when computers learn from data instead of being programmed with exact rules. They find patterns and improve over time.

AI can classify text as positive, negative, or neutral (sentiment analysis). It can also detect spam, categorize news, and sort support tickets.

During training, AI makes predictions, checks them against correct answers, calculates the error, and adjusts its internal settings to reduce mistakes.

Datasets come from many sources: web scraping, user contributions, sensors, and research. The quality and ethics of data collection matter greatly.

Raw data often needs cleaning — removing errors, filling gaps, and standardizing formats. Clean data helps AI learn more accurately.

If training data over-represents one group, AI becomes biased. A facial recognition system trained mostly on light-skinned faces works worse for darker skin.

Data augmentation creates new examples by modifying existing ones — flipping images, adding noise, or changing brightness. This helps AI learn with less data.

Instead of examples, RL uses rewards and penalties. An AI agent tries actions, gets rewards for good ones, and learns to maximize its score.

Email spam filters, Netflix recommendations, voice assistants, and auto-tag on Facebook all use machine learning trained on specific datasets.

Accuracy measures how often AI is right. Precision and recall measure different types of errors. Choosing the right metric depends on the task.

Collecting data responsibly means getting consent, respecting privacy, and being transparent. AI trained on stolen or private data raises serious ethical issues.

Training data is the collection of examples AI uses to learn. For a cat recognizer, training data would be thousands of labeled cat photos.

ML trains on data to find patterns. Understanding training, testing, bias, and evaluation helps us build and use AI responsibly.

In supervised learning, every example has a correct answer (label). AI learns by comparing its guesses to the right answers and adjusting.

In unsupervised learning, data has no labels. AI finds patterns and groups on its own, like clustering similar customers or topics.

Features are the characteristics AI looks at (like pixel colors in an image). Labels are the correct answers (like 'cat' or 'dog').

Data is split into training data (to learn from) and test data (to evaluate). Testing on new data shows if AI actually learned or just memorized.

Overfitting happens when AI memorizes training data too well and fails on new data. It is like memorizing answers instead of understanding concepts.

Underfitting happens when AI has not learned enough and performs badly on everything. It needs more data or a better model architecture.

Image classifiers sort pictures into categories. AI learns from labeled photos to classify new images as dogs, cats, cars, or thousands of other objects.