Project 4 · workbench in progress

A perception system under stress

Train a model, choose on development evidence, then freeze a test and explain its failures. This saved synthetic workbench is available now. Separate objective cases are available below. The P06 real-image baseline is runnable; its fresh-source evaluation and independent review remain outstanding. No project-completion award is available yet.

Convolution refresher · Augmentation lesson · Real-image annotation practice · My Work and backups

Loading local saved work. Controls will become available when initialization finishes.

Choose your evidence

Project variant

Predict a failure, inspect an image and its model trace, and explain a frozen comparison. You can change variants without clearing work.

1. Train and select on development evidence

5 × 5 pixels→Two learned 3 × 3 kernels
18 weights + 2 biases→Tanh maps → spatial means→Logistic head
3 parameters

All 23 parameters learn from actual cross-entropy gradients. Vertical = 0, horizontal = 1. Convolution uses valid windows with no padding; averaging each 3 × 3 feature map produces two features. The model does not see development or reserved rows while training.

Training recipe — changing coverage or rate restarts the candidate

Training coverageLearning rateHorizontal probability thresholdRequired accuracy in every group

0/1,000 updates. Jobs cap at 25; requests cap at 250 and commit only when all jobs finish. Canceling, hiding, scrolling away from this section or navigating away retains the preceding model. Direct CPU controls do not require a worker or WebGPU.

Development: 17/36 correct. Always-vertical baseline: 18/36. These 36 inputs are two related views of eighteen scenes, repeatedly available for selection.

clean: 8/18; baseline 9/18.
background: 9/18; baseline 9/18.

Training uses foreground brightness 0.5/1 on black backgrounds. Development uses brightness 0.75 on black/0.1 backgrounds. Inputs differ, but generating rules and positions overlap. Good development scores do not prove real-image transfer.

Selection reason (1,200 characters)

2. Reveal a frozen stress test

Known exposure: not yet recorded locally. These are public synthetic fixtures, not a secret or authenticated independent assessment. Three stress views share each of eighteen scene identities; 54 views are not 54 independent scenes.

0/4 captures retained. Resetting the candidate or importing an older backup cannot erase known exposure or the first local result.

A failed gate is useful evidence, not a reason to hide a run. A passing gate applies only to these fixed synthetic groups.

3. Inspect a failure and its mechanism

Gallery evidence Show only errors

Showing the current model on repeated development examples. 36 matching examples.

Example

Input and decision

Target vertical; predicted vertical. Horizontal probability 49.65%. Scene development-0-1-1.

Channel 1

Mean 0.0844. Common shading scale: −1 transparent to +1 navy. Values: 0.199, 0.170, 0.000, 0.058, 0.233, 0.000, -0.047, 0.147, 0.000.

Channel 2

Mean 0.0338. Common shading scale: −1 transparent to +1 navy. Values: 0.207, -0.224, 0.000, 0.246, -0.092, 0.000, 0.140, 0.027, 0.000.

Head score -0.0141 becomes horizontal probability through the logistic function. A brighter background changes many local windows even when orientation is unchanged.

4. Compare with real-image evidence

Use the eight licensed Oxford-IIIT Pet images to expose annotation ambiguity and pixel-model failures. This is repeated real-image practice, separate from the synthetic frozen test. It does not supply a reserved real-image evaluation or prove that the bar classifier transfers to photographs.

0/4 real-image captures.

The saved mask and model stay separate from source labels. The derived foreground-envelope box is not the dataset's animal-head annotation. A low overlap is valid failure evidence. Dataset attribution and transformation record. Human review and learner pilots remain pending.

5. Write and export the evidence

Saved report — 2,000 characters per fieldData and split card

Describe the authored images, related views, training coverage and what real-world data is still missing.

Representation diagram and explanation

Explain 25 pixels → two learned 3×3 kernels → tanh → two spatial means → a probability. What does global averaging discard?

Baseline and frozen comparison

Compare with always predicting vertical. State the chosen threshold, subgroup floor and why you froze this candidate.

Failure gallery and intervention

Name a captured example ID, inspect its probability and feature maps, and propose a controlled label-preserving intervention.

Recommendation

Explain whether the evidence supports use, further experiments or rejection. A failed model can support a strong decision report.

Limitations and next independent check

State prior test exposure, shared synthetic generating rules and missing real-data/independent evidence.

Use licensed pet-image practice to compare these toy inputs with real annotation ambiguity. Save selected real-image evidence above; the full export recomputes it from the pinned source. The P06 studio supplies a runnable pixel baseline; its independent extension, fresh-source evaluation and human review remain pending. Reports are not graded for reasoning or code quality. Learner pilots remain pending.

Separate objective transfer assessment

Try a different problem independently

These three fixed cases use different data and separate scoring arithmetic from the guided perception project. Work without the walkthrough first. Opening and submissions are saved; the first submission is retained, with at most two practice retries. Reopening, resetting or importing an earlier backup does not erase known exposure. This is a local objective check, not authenticated exam supervision or human review of your implementation.

0/3 cases passed on first submission. Latest/practice passes: 0/3. Project evidence, these outcomes and pending human review are separate.