Why Broadcasting Is NumPy's Superpower (and Its Silent Bug)

One missing dimension can wreck your model, and NumPy will never warn you.

• No error raised
• Wrong answer anyway

Meet broadcasting, NumPy's quiet superpower.

• NumPy's quiet superpower

Take a column of three numbers, and a row of four. Add them. NumPy stretches both into a three by four grid. No loops.

• (3,1) plus (1,4)
• Result is (3,4)
• No loops, no copies

Here is the one rule. Compare shapes from the right. Each pair must be equal, or one of them must be one.

• Compare from the right
• Equal, or one

Three by one against one by four? Four versus one works. Three versus one works. Done.

• 1 versus 4: works
• 3 versus 1: works

Now the silent bug. Predictions shaped n, labels shaped n by one. Subtract them, and you get no error. You get an n by n grid, and a wrong loss.

• (n,) minus (n,1)
• Gives (n,n), not (n,)
• Loss is quietly wrong

The fix is simple. Reshape so both are n by one, and always print your shapes.

• Reshape to (n,1)
• Always print shapes

Lesson thirty-one builds NumPy arrays from the ground up. Follow along.

• NumPy arrays from scratch