Little Bits of Artificial Intelligence

Q1: Why does it add a constant (bias/intercept)?

Imagine we’re fitting a line through dots.

Without a constant:

If we only use y = m * x, our line would always go through (0,0). But what if our data looks like this?

Clearly, y = 2x + 1 fits best — not y = 2x.

That +1 is the bias (or intercept) — it lets the line shift up or down to match the data.

Think of it like this:

• The slope (m) tilts the line.
• The bias (c) moves the line up or down.
• So: it adds the constant to better fit data that doesn’t start at 0.

Q2: When does it adjust the slope and the constant?

Answer: Every training step!

During training (every loop in gradient descent), the model:

1. Sees the prediction is wrong (error)
2. Figures out how wrong it is
3. Calculates how much to adjust both the slope (m) and bias (c)
4. Updates them to reduce the error

This happens on every iteration (called an “epoch”) until the error becomes really small.

Q3: How does it decide how to adjust them?

It uses gradients (think: slopes of the error curve).
Let’s visualize it simply:

• The model wants to find the lowest error.
• It draws a little hill for error vs. weight or bias.
• Then it walks downhill toward less error using calculus.

The math (we saw this in the gradient descent part):

m -= learning_rate * total_error_m
c -= learning_rate * total_error_c

If the error says: “Your guess is too low,” it increases m or c

If the guess is too high, it decreases m or c

This continues until the guesses are very close to the actual values

Summary: One-Liner Answers