Preston AI Lab
Preston AI Lab / Optical studyIllustrative halftone print

Shape and signal.

Proposal for Preston AI Lab · Runs in your browser · No trained model

Could you still recognise a Loop if the ink broke up? Shape and Signal is a design proposal for Preston AI Lab. It helps students and teachers of visual perception compare how a person and a simple rule read a mark. Pick a mark, predict what the rule will call it, then see which pixels changed its answer.

Open observation desk

Observation desk

Start with a clean impression. Keep the source selected while you change one control at a time. The source tells you what went in. The three distances tell you which shapes the remaining ink resembles.

Source glyph

Observed pixels

Which template do you think will win?

5 columns × 7 rows

Inspect a cell

Shade 0 of 255

Closest template

Arch

Mean difference

A smaller difference means a closer match. Zero is a pixel-for-pixel match.

From the observation notebook

Two readings of a Loop.

Imagine a design class preparing a sheet of tiny printed symbols. Someone spots a smudged Loop and asks how much of its bottom stroke has to survive. Before looking at a scan, the class can try that question here under repeatable conditions. These two impressions use the same source and keep blur at zero.

Impression A0% noise
Rows 6 and 7 · The closing stroke

The closing stroke

The clean Loop matches all 35 cells of its own template. Arch has exactly the same first five rows. The last two rows differ in seven places, so the distance to Arch is 7 divided by 35, or 20%. Cover the bottom of the observed grid and see how much of the picture could belong to either glyph.

Arch
20.00%
Loop
0.00%
Fork
71.43%
Impression B75% noise
Row 2, column 2 · Shade 226

A Loop read as Fork

At this setting the fixed noise pattern replaces 29 of the 35 cells. Look at row 2, column 2. Loop expects black (0), Fork expects white (255), and the observed shade is 226. That cell contributes 226 units of error to Loop and only 29 to Fork. Across all 35 cells the errors add up to 4,556 for Loop and 3,809 for Fork. The reader calls it Fork even though you supplied a Loop.

Arch
54.22%
Loop
51.05%
Fork
42.68%
Method notes

Reading a mark.

This proposed study uses three synthetic templates drawn on the same grid. That constraint lets a class inspect every pixel behind a result. Everything runs in the browser using a fixed matching rule.

No model was trained.

For a class record, write down the source, both slider values, your own reading and all three distances. Give just the settings to a partner. They can reconstruct the impression and compare their reading with yours before checking the scores.

  1. 01 / Draw
    Keep the paper still.

    Each glyph occupies five columns and seven rows. Black is 0 and white is 255. Its size, position and orientation stay fixed in every impression. The desk accepts these three sources. A photograph of a sign introduces changes of scale and angle that this study leaves open.

  2. 02 / Vary
    Change one thing at a time.

    Blur mixes a cell with the average of its surrounding patch, including the cell itself. At an edge the patch ends at the grid. Noise replaces selected cells with fixed shades. Its slider sets a threshold, so 75% need not replace exactly three quarters of a small grid. The same settings always give the same impression.

  3. 03 / Read
    Give every cell one vote.

    Take the absolute difference between each observed shade and its partner in a template. Add the 35 differences and divide by 8,925, the largest possible total, to get a fraction. The desk displays it as a percentage. The smallest unrounded total wins, and equal lowest totals share the result. A score of 42.68% describes a distance between pictures.

Classroom record

Now write the record.

Choose a source, set blur and noise, and write down your own reading before you look at the scores. Then note all three distances. Give only the settings to a partner. They can rebuild the same impression and compare their reading with yours, because the same settings always give the same picture.

This independent design proposal has no published team contact. The record is yours to use in class.

Observation slip

Source
Arch
Blur
0%
Noise
0%
My reading
Arch
0.00%
Loop
20.00%
Fork
68.57%
Save this record (.txt) ↓