7AAVCD42 Chap.2 Machine Vision and Model Opacity
Machine Vision and Model Opacity
From image elements to institutional decisions
Machine vision begins with numerical image elements and builds learned features through network layers. This chapter keeps that pipeline visible so that words such as seeing and recognising do not imply human perception. It distinguishes technical opacity inside a model from organisational opacity around data, deployment and contestability.
The method helps you locate an error, request the right evidence and explain how an output becomes consequential when an institution acts on it.
Reconstruct the representation pipeline
Begin with the image available to the system, not the object a person thinks the image shows. Record the numerical elements and preprocessing conditions.
Explain how training makes recurring patterns useful as features and how later layers combine those features into an output. A correct label does not prove that the system used a humanly meaningful feature, and an incorrect label may begin with image quality, training coverage or the relation between features.
Use changed conditions to diagnose the pipeline.
Blur, lighting, partial occlusion, camera angle or background can alter the represented input. Ask whether comparable conditions appeared in training and evaluation. If performance falls, locate the earliest stage affected before proposing a remedy.
More data is not a complete answer unless the missing condition and labelling practice are identified.
Name the kind of opacity
Internal opacity concerns difficulty explaining how model relations produced one result. Data opacity concerns missing knowledge about collection, labels and coverage. Institutional opacity concerns thresholds, use rules, human review and appeal.
These problems can coexist, but they require different evidence. A feature visualisation cannot reveal a procurement decision, while a policy document cannot show whether the model relied on an unstable image region.
Replace the phrase black box with a precise request.
Ask for evaluation across relevant image conditions, group-sensitive error evidence, threshold policy, documentation of reference data or a local explanation for a particular output.
A reviewer can then assess whether the missing information is technically unavailable, commercially withheld or simply not collected.
Connect error to consequence
A model output becomes socially important when an institution assigns an action to it. Identify whether the output unlocks a device, ranks a case, alerts an officer or supports another decision.
State what a human sees and whether they can override the result. The same classification technology can carry very different stakes and power relations across settings.
In coursework, organise a case into input, representation, output, institutional use and remedy.
Give confidence and accuracy separate roles: confidence ranks outputs under the model, while accuracy requires comparison with correct outcomes under suitable conditions. Finish by naming the evidence that would change your diagnosis and the actor able to repair the decision process.
What this chapter covers
- 01
Image Elements
- 02
Learned Features
- 03
Neural Networks
- 04
Model Opacity
- 05
Accountability
Worked example · free
Why confidence can mislead
- 2Inspect the input conditions.
- 2Trace learned features.
- 2Separate confidence from correctness.
Key terms
- Learned Feature
- Learned Feature is a pattern derived from training data and used by later network layers.
- Model Confidence
- Model Confidence ranks outputs within a model and does not independently guarantee factual correctness.
- Local Explanation
- Local Explanation describes which represented inputs influenced one particular model output.
Machine Vision and Model Opacity FAQ
How does a model see an image?
It receives numerical image elements and transforms them through learned features and layers. The output is a model representation and classification, not direct human-like perception of the original object.
What kinds of opacity should I distinguish?
Separate difficulty explaining an internal prediction from missing information about training data, thresholds, deployment rules or appeal. Each form of opacity calls for different evidence and remedies.
Why is confidence not the same as accuracy?
Confidence expresses the model's ranking among its available outputs for that input. Accuracy is established through comparison with correct outcomes across suitable external evaluation conditions.
Assessment move
Draw the image-to-output pipeline, mark one possible failure at each stage and add the institutional decision that follows the output. Change one image condition, such as blur, lighting or occlusion, and predict which represented features will weaken before checking the model result. Make a second version of the diagram for organisational opacity: training documentation, threshold policy, reviewer instructions and appeal.
Practise explaining why confidence and correctness answer different questions. Then choose two deployments that share a technical comparison step and list how consent, stakes and remedies differ. Your final retrieval sentence should name the missing evidence precisely, because a request for evaluation data or threshold documentation is more useful than calling the whole system a black box.
End the session by explaining one false classification twice: first to a technical reviewer and then to a person affected by the institutional decision. Notice which evidence and vocabulary each audience needs. Sketch the pipeline again the next day and repair the first missing stage.