Perjantai — past, present, future

Is the future earning its keep?

skill vs copy-last-frame

—

1 − error(model) / error(just repeat the last frame)

model error

—

copy-last error

—

trust in the eye (K)

—

noise floor —

belief vs eye

—

frames · rate

—

duplicate frames skipped: 0

The older past: 20 learned weights

now 3×3

motion 3×3

mem · bias · phase ×2

Tenses

Past moon memory

How long the past moon remembers. Long memory lets a still room be predicted from the past instead of the noisy present.

Plasticity

How fast the 20 weights forget old evidence.

Camera ↔ dream

Below 100% the brain is fed its own prediction. Learning pauses: a dream must not teach itself. Watch fine detail dissolve, then pull it back to 100%.

Future moon horizon

Frames ahead. The model is rolled forward from the present; further out, the future blurs.

Surprise gain

Precision: the present is a beliefPresent = prediction + K·(eye − prediction). K is low where the room is predictable, so the present moon shows a denoised belief. Anything new breaks through at once. Phase moon (experimental)Remembers phase velocity in a complex Gabor bank and offers it to the predictor. Measured: +1.7 pts on the room, +3–4 pts only on fast motion, killed on its own pre-registered test. Costs ~8 ms/frame. LearningOff: the weights freeze where they are. Future moon goes where surprise livesOff: it parks in the middle. Attention gates learningThe weights learn mostly from where the future moon is looking.

What is real here

Past is two things: the past moon's slow average of what was seen, and the 20 weights, which are the summary of everything learned.
Present is the frame that just arrived, 64×64.
Future is the past read through the present: p = w · φ(x, x − x₋₁, mem). One linear filter shared by every pixel, solved by exact least squares with forgetting.
Walls show |camera − prediction|, the innovation, above the frame's own noise floor (3×3 average minus the frame mean). The innovation is the only signal that teaches.
Precision (on by default): the present moon shows a belief, p + K·(eye − p), with per-pixel K = 1 − R / max(running surprise, this frame's local surprise). On a still noisy room the belief is 7.5× closer to the clean scene than the raw camera, and a sudden new object passes through in its first frame (8% error).
Phase moon (off by default): its pre-registered test was killed, since the linear filter already gets 98.7% on smooth slow translation. It earns +3–4 points only on fast motion. Kept as an experiment.
The future moon moves to the peak of running squared surprise. That is the Kuulustelu rule: look where you expect the answer to split. It is programmed, not learned.
This is not a deep video model. It is the smallest model that has to beat copy-last-frame, and the skill meter says honestly whether it does. In tests it beats it by ~47% on a still noisy room and ~91% on smooth noiseless motion; real cameras vary.
Measured surprise: the dream does not melt to grey. The learned filter keeps the overall brightness, so the layout persists while fine detail dissolves.
Your camera frames never leave this page.