Data & Robots Lab

Kit

A chart answers one question, so I always ask what the question is before I draw. Change over time wants a line. Comparing a few groups wants bars. Parts of one whole can be a pie, as long as there are only a few parts. Two measured numbers that might move together want a scatter plot.

Kit

Cô Lan's oven camera gives every loaf a burnt score. I have to pick a line, and every loaf above it goes in the bin. Put the line low and I catch all the burnt ones but bin good bread too. Put it high and nothing good is wasted, but burnt loaves get sold. Precision asks: of the loaves I binned, how many were really burnt? Recall asks: of the burnt loaves, how many did I catch?

Kit

Here is one busy morning at the cart, sales counted every 10 minutes. I learn a curve from the solid dots, then check it on the hollow dots I never saw. A bendier curve (a higher degree) always fits the solid dots better. Past a point it starts copying the noise, and the error on the hidden dots shoots up: it has overfitted.

Kit

My delivery drone has to hover at 10 metres. P pushes harder the further off I am, but on its own it overshoots and swings. D watches how fast I am moving and brakes, which calms the swing. With only P and D, gravity still holds me a little low, so I keeps a running total of that leftover error and pushes it away.

Kit

A scooter's GPS jumps around by several metres. I also know roughly how fast it is going and which way. A Kalman filter blends the two: guess where the scooter should be, look at the GPS, and move part of the way toward it. How far depends on which one I trust more. Robots and phones use the extended version, the EKF, when the motion curves; the loop is the same.

Kit

My robot needs a way from the cart to a customer. Dijkstra searches in rings around the start: it never misses, but it checks a lot. Greedy rushes at whatever looks closest to the goal and walls can fool it. A* adds the distance already walked to a guess of what is left, so it heads for the goal and still finds a shortest path.

Kit

Training a model means walking downhill on its error. I feel the slope where I stand and take a step down. The learning rate is the size of that step. Too small and I am still walking at dinner time. Too big and I jump right over the valley, each jump wilder than the one before.