PYTHON · NUMPY

A kilobyte classifier

Train a tiny softmax classifier on features -- on-device learning, the TinyML rung -- and see it generalize to views it never saw.

01Challenge

You have features; now learn from them. The Bench gives three classes of feature clouds and a softmax classifier with its weights stuck at zero -- the gradient is already computed for you, but nothing updates. Take the gradient-descent step so the classifier learns, and watch it label a held-out test set it never trained on. This is TinyML: a kilobyte of weights, trained in place, exactly what a $16 Grove Vision or a reCamera does on-device.

Predict firstBefore any training, the classifier's weights are all zero. What accuracy will it get on 3 classes?

02Model

A softmax classifier turns features into class probabilities; cross-entropy measures how wrong they are; the gradient G = P - onehot(y) points downhill. One line -- W -= lr*gW -- is the entire act of learning. That the trained classifier scores high on data it never saw is generalization, the thing that separates learning from memorizing.

03Guided practice

Step A (worked): the gradient gW, gb is computed each step, but the update line is a bare pass, so the weights stay zero and test accuracy sits at chance (~33%) -- the FAIL. Step B (the fix): W -= lrgW; b -= lrgb. Step C: accuracy climbs past 90% on the held-out set. Step D (connect): this is the on-device 'Teach it to see' lab in numbers -- collect, train in place, infer.

04Feedback

PASS WHEN PASS when the trained classifier reaches at least 90% on the held-out test set.

If your run fails, check:

Read the bench’s printed result — it passes when the trained classifier reaches at least 90% on the held-out test set. Run the cells top to bottom; if one errors, the FAIL line names the step to fix.
Accuracy is still near chance -- the weights never update; add W -= lr*gW and b -= lr*gb.

05Retrieve & space

What does the gradient P - onehot(y) tell the weights to do?
Why score on a held-out test set instead of the training data?
What makes this 'TinyML' -- what is small about it?

06Mastery & project

Pass the lab: a classifier you trained from scratch labels views it never saw. That is on-device learning -- the second rung.

This is the numbers behind the live 'Teach it to see' lab, and the heavy-perception path (real detection and depth) runs the same idea at scale.

GO FURTHER · THE LIVE PLAYGROUND

Do it live: Teach it to see ↗ — train this classifier from your webcam, on-device.
Heavy perception, on-device ↗ — real detection, depth, and segmentation in a browser tab.

← The sensor speaks in numbers Perception becomes a decision →

PYTHON · NUMPY

A kilobyte classifier

Train a tiny softmax classifier on features -- on-device learning, the TinyML rung -- and see it generalize to views it never saw.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you