Does KNN need training?

No weight learning—predictions come from the majority label among the K nearest training points. Change the query and neighbors update instantly.

What happens when K is large?

Small K yields jagged boundaries sensitive to noise. Large K smooths boundaries but can misclassify mixed regions.

How is this different from the neural network classifier?

Same 2D data, but KNN uses only distance and majority vote while the NN classifier learns weights via backprop. Compare both playgrounds to feel classical ML vs deep learning.

NN Classifier Loss Landscape Lab KNN Neighbor Classifier RL Agent Conv Vision Attention Playground

Playground

KNN Neighbor Classifier

Classify with a majority vote from the K nearest neighbors—no training step! Compare classical ML on the same K-datasets as the neural network classifier.

Neighbors KDistanceShow formulas

Control help

Click the canvas to move the query (purple ring); dashed lines connect the K nearest neighbors.

Neighbors K: How many nearest training points to use. Small K = jagged; large K = smoother.
Distance: How to measure distance between points—same L1/L2 ideas as in ml03.

Data

Which K-culture dataset?

Train/Test ratio 70%Noise 0.5

Show test data

Classification

Background = KNN decision region · dashed = neighbor links · click = query

Click the canvas to move the query and inspect neighbors

Test accuracy

0.8167

Query prediction

Negative (−1)

K-vote breakdown

+1 / −4

#1 · d=0.21
#2 · d=0.33
#3 · d=0.51
#4 · d=0.67
#5 · d=0.68

Related lessons