PYTHON · NUMPY

The sensor speaks in numbers

You feed raw 8-bit sensor readings, values from 0 to 255, straight into a learner. What goes wrong?

Turn a stream of raw 8-bit sensor readings into a small, centred feature vector a model can learn from.

01Challenge

A camera or a sensor array gives you raw bytes: integers from 0 to 255. A learner wants small numbers centred near zero. The Bench hands you three raw frames and a featurize() that returns them untouched; scale and centre them into [-0.5, 0.5]. This one move -- raw reading to feature -- is the first rung of every perception system, and it is exactly what runs first on a Grove Vision module before any model sees the data.

Predict firstYou feed raw 8-bit sensor readings, values from 0 to 255, straight into a learner. What goes wrong?

02Model

Neural nets learn fastest when inputs are small and centred; raw 0-255 bytes swamp the early gradients. Dividing by 255 puts values in [0,1]; subtracting 0.5 centres them. That is the whole preprocessing step, and it is the same line the on-device 'Teach it to see' lab runs on every camera frame.

03Guided practice

Step A (worked): featurize returns the frame unchanged, so features span 5 to 255 -- run it and read the FAIL. Step B (the fix): return frame / 255.0 - 0.5. Step C (check): the dark frame becomes small negatives, the bright frame small positives, all inside [-0.5, 0.5]. Step D (feel it): a model trained on centred features converges in a fraction of the steps of one trained on raw bytes.

04Feedback

PASS WHEN PASS when featurize scales and centres every reading into [-0.5, 0.5].

If your run fails, check:

Read the bench’s printed result, it passes when featurize scales and centres every reading into [-0.5, 0.5]. Run the cells top to bottom; if one errors, the FAIL line names the step to fix.
The features are still outside [-0.5, 0.5] -- divide by 255, then subtract 0.5.

05Retrieve & space

Why centre features near zero before training a model?
What range does frame / 255 - 0.5 produce?
Where does this preprocessing run on a real vision module -- before or after the model?

06Mastery & project

Pass the lab: raw bytes become centred features. You have built the first rung -- the perception primitive every model sits on.

Feeds the next lab, where these features train a real classifier, and mirrors the live 'Teach it to see' lab that featurizes your webcam on-device.

GO FURTHER · THE LIVE PLAYGROUND

Feel it live: Teach it to see ↗ · the same featurize step, running on your webcam, on-device.

← Perception to Policy: From a Sensor to a VLA A kilobyte classifier →

PYTHON · NUMPY

The sensor speaks in numbers

Turn a stream of raw 8-bit sensor readings into a small, centred feature vector a model can learn from.

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