@dailymathvisuals: Last episode, the line failed. So today, a question. What if we let the line bend? One bend. Two bends. Three. Five. Fourteen. So where do we stop? That's the whole story. Imagine a straight line with joints. No joints — just a line. One joint — now it can bend. More joints, more freedom. The curve has enough flexibility to follow the data more closely. Those joints have names. p of x equals a zero plus a one x plus a two x squared, and so on. Polynomial regression. Same loss as linear regression. Minimize the sum of squared errors. In fact, despite the name, polynomial regression is still linear regression — because the model is linear in the coefficients. Only the input features have changed. So what does each term actually do? A constant gives you a horizontal line, no bend. Add the linear term — now it tilts. Add x squared — this is the first polynomial that can curve, a parabola. Add x cubed — a cubic can introduce an inflection point, allowing much richer shapes. Each new term, one more way to bend. But more flexibility isn't always better. Imagine there's a true smooth pattern in the world. You only see noisy samples from it. A high-degree polynomial has enough flexibility to closely match every noisy point. But on a brand-new point we never trained on, it misses completely. That's overfitting. So how do you pick the right degree? Watch the errors. Train error keeps falling. But test error typically drops at first, then climbs back up. The sweet spot is the test minimum. Every extra degree gives the model another way to bend. Too few, and it can't see the pattern. Too many, and it sees patterns that don't exist. Machine learning is the art of stopping at exactly the right amount of model flexibility. Next, k-Nearest Neighbor. Follow at daily math visuals for the full series. Shop: dailymathvisuals.lemonsqueezy.com #math #machinelearning #regression #polynomialregression #overfitting #biasvariance #datascience #ml #ai #mathtok #stem

Dailymathvisuals

Open In TikTok:

Region: US

Monday 29 June 2026 04:28:27 GMT