DSAverse

Initializing Sorting Algorithms...

Sorting

Trees

Graphs

Preparing interactive visualizations...

Back to Machine Learning

What Is Machine Learning?

Fit a line to twelve points by hand, then let gradient descent do it. Learning is a search: for the knob settings that make one number, the loss, as small as it can be.

Builds onGradient Descent Least Squares

Fitted live in your browser · seeded example data

Act 1 of 6 · Examplesstep 1 of 22

● examples— the model's line

w (tilt) -1.5 → 2.5 · b (height) -2 ↑ 4× = lowest loss

Speed

Twelve examples, each an input x and the answer y that came with it. Nobody wrote down the rule that turns x into y. Machine learning is finding a rule like that from the examples alone, so it can answer an x it has never seen.

\{(x_i,\, y_i)\}_{i=1}^{12}, \qquad (x_1, y_1) = (-3.3,\, -1.4)

Every learner has three pieces

A model family — rules with knobs: here every line ŷ = w·x + b.
A loss — one number for how wrong a setting is: the mean squared error.
A search — a way to find low-loss settings: trial and error, or gradient descent.

A common misconception

“Train long enough and the computer will work out the right answer.”

It only ever searches the family it was given. On a set of curved examples the best straight line has loss 2.612, and gradient descent reaches exactly that and stops — a million more steps change nothing. A parabola gets to 0.254. Choosing the family is a human decision; try “Curved examples” in Explore.

Check yourself

Question 1 of 3

What does “learning” mean for a model like ŷ = w·x + b?

What Is Machine Learning?

Fit a line to twelve points by hand, then let gradient descent do it. Learning is a search: for the knob settings that make one number, the loss, as small as it can be.

Builds onGradient Descent Least Squares

Fitted live in your browser · seeded example data

Act 1 of 6 · Examplesstep 1 of 22

● examples— the model's line

w (tilt) -1.5 → 2.5 · b (height) -2 ↑ 4× = lowest loss

Speed

\{(x_i,\, y_i)\}_{i=1}^{12}, \qquad (x_1, y_1) = (-3.3,\, -1.4)

Every learner has three pieces

A model family — rules with knobs: here every line ŷ = w·x + b.
A loss — one number for how wrong a setting is: the mean squared error.
A search — a way to find low-loss settings: trial and error, or gradient descent.

A common misconception

“Train long enough and the computer will work out the right answer.”

Check yourself

Question 1 of 3

What does “learning” mean for a model like ŷ = w·x + b?

DSAverse

What Is Machine Learning?

See also

What Is Machine Learning?

See also