AI & ML: How Machines Learn

Site Admin · 11 Sep 2026 · 6 views

The Learning Loop

A machine learning project is an ordinary loop repeated until the model is good enough. You gather data, prepare it, train a model, measure how wrong it is, adjust, and repeat. The measure of wrongness is the loss, and training iteratively lowers it.

Features, Labels, and Samples

Rows in a dataset are samples, and columns are features. The label is the answer the model should predict. In a house price dataset, square footage and location are features, and the selling price is the label. The model reads the features and learns a function that roughly turns them into the label.

# one training sample
# features: size_sqft, bedrooms, age_years
# label: price_usd
(1200, 3, 15) -> 245000

The arrow means predict: given those features, output around 245000. Many such examples teach the model the relationship.

Training Means Adjusting Weights

Inside many models, the learned knowledge lives in numeric weights. Training starts with random weights, makes a prediction, compares it to the true label, and nudges the weights to reduce the error. The nudging rule is gradient descent, and how far each step moves is the learning rate.

prediction = features * weights
loss = (prediction - label)^2
weights = weights - learning_rate * slope_of_loss

These three lines capture the heart of gradient based learning: a prediction, a loss that punishes mistakes, and a weight update that walks downhill on the error surface.

Underfitting and Overfitting

An underfit model is too simple and misses the real pattern. An overfit model memorises the training data, including its noise, and fails on new examples. The solution is usually more or better data, simpler models, and always checking performance on data the model never saw during training.

Data Quality Beats Math

Garbage in produces garbage out. Duplicated rows bias statistics, missing values break training, and skewed classes hide poor performance. Most professional time is spent cleaning data, defining features, and designing solid evaluation, not inventing new algorithms.