An invented class. A neural network learning to recognise grade bands.
This tool trains a real neural network on an invented class of students, sorting them into
four grade bands. The class map shows every student by their true band — filled dots are
training students, hollow dots are unseen (held back to check what the network really
learned), and rings appear for a check-up group once early stopping is on. Its background is
the network's current guess, sliced through the two measurements you're looking at —
open "Adjust the typical / what-if student" to move that slice, or to build a hypothetical
student and see its prediction as a star on the map. Hover a dot for that student's own card,
or a unit in the network view to see what it responds to. Below the mistakes chart, a
confusion grid and an at-risk alert show how the network is doing and where it goes wrong.
Reveal how the grades were really made (Shift+R) when
you're ready to compare notes. Nothing here is a real person — names, measurements and
grades are all invented.
The steps are so big the network overshoots every time. Try a smaller step size, then start over.
The class map
Shapes show true grade bands. The background is the network’s guess for the two measurements shown.
Inside the network
Units, layers and weights. Hover a unit to see what it responds to on the map.
How it's going
Press Play to watch mistakes change over training rounds.
training
—
unseen
—
placed in the right band
Explore the mistakes Confusion grid & at-risk alert
Who does it confuse with whom?
caught
—
missed
—
false alarms
—
Should a tool like this be used on real students? Topic 10 comes back to this question.