Mario brain scope
A Cadence brain (U3ck83) playing Super Mario Bros. alone, no coach. Every number below was recorded from the real brain at each decision. Scrub, play, click neurons. Or watch W1b's whole World 1 gauntlet with its brain.
Loading...
decision 1 / x pressed
What it sees (its real input: 26 x 32 grey blocks, 0 = background, shown 3x brighter)
now
previous decision
Body sense (42 numbers from Mario's body: x speed, y speed, ground/jump/fall/pole, rising, facing, power, hurt/star, time left, held A B right left)
Neurons (final activation after settling; click one, or a region name, for its trace)
association
sorted by the button each one pushes hardest
prefrontal
working memory; silent, since predict() reads no live memory
sensory
eyes now | eyes previous | below: body sense, 8 copies
Buttons over the whole run (7 motor neurons' activation per decision; white line = the pressed one, green = x progress; click or drag to jump)
Buttons now (motor neuron activation; the highest one is pressed)
Settle of this decision (7 motor neurons, step by step)
Trace:
Activation over the whole run, cursor at the current decision. Click the trace to jump.
Brain-cam videos
The four runs in the scope as videos at game speed: game on the left, the recorded brain (U3ck83) on the right.
What you're looking at
The brain is mario/brains/U3ck83.npz: 2,519 neurons in four regions of its connectome (sensory 2,000, association 256, prefrontal 256, motor 7). The same weights clear World 1-1, 1-3 and 1-4. Each run here is a real solo attempt from a fixed start, replayed on the real NES emulator.
- Decision: every 4 NES frames the brain gets one observation and settles. The most active motor neuron after settling is the button it holds for the next 4 frames.
- Neuron colors: blue is inhibited (about -0.1), black is 0, red to yellow is up to 1.0. Activity is stored at 8 bits, so values are within about 0.002 of the recording.
- Prefrontal is dark on purpose. In this mode the brain answers each screen on its own (Cadence
predict, no live memory), so the working-memory region never gets input. That's the real state, not a bug in the page. - Recording: made with
cadence.recording.record_settlements, which runs Cadence's NumPy path. The recorded choice matched the button actually pressed at every decision. - Poke the brain: for each decision, 4 x 4 block patches of the input (both frames) were blacked out one at a time, 180 patches, and the brain was settled again offline. Red means hiding that spot lowered the activation of the button it pressed (colors scaled per decision; the page prints the biggest drop). The count says how many patches made it press something else. One extra run zeroed the whole body sense.
Code and regeneration steps: mario/viz/ in the cadence-arcade repo.