Mei Tanaka

Day in the Life

Mei Tanaka’s screen is full of drone footage — thousands of aerial images of collapsed buildings, flooded streets, forested hillsides after storms. None of it is from a real disaster. Yet. It’s training data, carefully labeled: this pixel is a road, this pixel is a roof, this pixel is a person.

Her job today: build a model that can do, in half a second, what used to take a trained analyst minutes — scan a fresh drone photo and flag every spot that might be a person who needs help. Too cautious, it misses someone. Too eager, it buries rescue teams in false alarms, and every false alarm costs time.

She kicks off a training run and watches a graph update live: the model’s loss — how wrong it is, right now — ticking downward as it sees example after example and nudges itself, a tiny bit, each time.

By lunch, the model has seen the training images hundreds of times. Its error rate on photos it’s never seen drops from 38% to 6%.

But Mei doesn’t celebrate the number — she goes hunting for where it’s still wrong. She finds the pattern fast: people partially buried, shot from directly overhead, in low light. So this afternoon isn’t more training. It’s walking over to the data team and asking for exactly those kinds of images. A model is only as good as what it’s been shown, and Mei’s real skill is knowing what to show it next.

By evening, the updated model runs on footage from a real wildfire response last year. On a hillside photo full of ash and smoke, it flags two people the original human review missed. Mei doesn’t say much. She just forwards the result to the search-and-rescue team: “ready for field testing whenever you are.”

The Science

What a neural network actually is

A neural network is a mathematical function with a huge number of adjustable numbers (called weights and biases) — sometimes millions or billions of them. You feed it an input (like the pixels of a photo), it does a long chain of multiplications, additions, and simple nonlinear “squashing” operations, and out comes an output (like “person: 87% likely, at this location”).

Training = minimizing error with derivatives

A freshly created network is useless — its weights start out essentially random. Training is the process of nudging every one of those weights, a tiny bit, in the direction that reduces the model’s error on its training examples. The error is measured by a loss function, and the “which direction reduces the loss” question is answered using derivatives — specifically, the gradient, which is just the slope of the loss with respect to each weight.

$$ w_{\text{new}} = w_{\text{old}} - \eta \frac{\partial L}{\partial w} $$

This is called gradient descent: repeatedly take a small step ($\eta$, the learning rate) downhill on the loss “landscape.” It’s the same core idea as Problem A’s calculus — a derivative tells you the instantaneous rate of change — just applied to a loss function with millions of dimensions instead of a position curve with one.

Overfitting — memorizing instead of learning

If a model is trained too long, or has too many adjustable weights for the amount of data it has, it can start overfitting: getting better and better at the exact training examples while getting worse on new ones — like a student who memorizes the answer key instead of learning the method. That’s why Mei always checks the model’s performance on images it has never seen, not just the ones it trained on — and why getting more, and more varied, real-world examples often matters more than training longer.

Curious how a smooth curve’s slope — the same idea behind the gradient above — is computed and visualized step by step? Problem A walks through exactly that, for a spacecraft’s position curve. Try Problem A →