VIRTUAL LABORATORY / DATA SCIENCE & COMPUTER VISION
Data Exploration & Image Classification
Investigate a small image-feature dataset, train a transparent k-nearest neighbours (k-NN) classifier, tune its confidence threshold and evaluate its predictions. Every result is calculated from the data shown here.
VISUAL CONCEPT MAP · VISION
Data Exploration & Image Classification
See how images become pixels, features and model predictions.
Input / evidence Processing Output / insight
INVESTIGATION CONTROLS
Learning goal
Discover how feature values influence a model's classification, and why accuracy alone does not tell the whole story.
Discover how feature values influence a model's classification, and why accuracy alone does not tell the whole story.
Required columns: brightness,roundness,texture,label. Feature values should be between 0 and 1. A stratified 75/25 split is used, keeping each class represented in training and testing where possible.
Teaching model: the built-in dataset is synthetic and uses simplified numerical image features. It is not a production vision model and does not identify real photographs.
EXPERIMENT WORKSPACE
Dataset explorer
Built-in synthetic dataset · 24 rowsRECORDS24
FEATURES3
CLASSES3
Feature distribution · average by class
Feature relationship · brightness vs roundness
Explore the class patterns. Features are scaled from 0 to 1 so their distances can be compared.
Dataset records
| # | Brightness | Roundness | Texture | Label | Split |
|---|
Image feature classification bench
k-NN · Euclidean distanceChoose a test sample
BRIGHTNESS0.78
ROUNDNESS0.82
TEXTURE0.68
Nearest training examples
MODEL DECISIONCAT · 100%
Select a test sample and run classification.
k-NN compares a new sample with the k closest labelled training examples. The most frequent neighbour label becomes the predicted class. Confidence is the fraction of neighbours voting for that class; it is not a guarantee of correctness.
Held-out test evaluation
ACCURACY—
MACRO PRECISION—
MACRO RECALL—
Confusion matrix · actual rows / predicted columns
Class-wise performance
If confidence is below your threshold, the model abstains and flags the sample for review. Review how this changes coverage and accepted-prediction accuracy.
TEST SAMPLES—
ACCEPTED—
REVIEW RATE—
Run the evaluation to calculate metrics from the held-out records.
Digital practical notebook
Record your investigation| Trial | k | Threshold | Selected sample | Prediction | Confidence |
|---|
Knowledge check
5 questions · 5 marksChoose one answer for each question and submit.
Practical challenge
Change k and the confidence threshold. Find a setting that reduces false positives without rejecting too many correct samples. Record the trade-off in your notebook.
Change k and the confidence threshold. Find a setting that reduces false positives without rejecting too many correct samples. Record the trade-off in your notebook.
Teacher assessment rubric · 10 marks
Data exploration (2) · Understanding k-NN (2) · Evaluation interpretation (2) · Experimental record (2) · Critical thinking and responsible AI (2).
Data exploration (2) · Understanding k-NN (2) · Evaluation interpretation (2) · Experimental record (2) · Critical thinking and responsible AI (2).
LIVE MODEL INSIGHTS
SELECTED CLASSCAT
TOP-CLASS CONFIDENCE1.00
MODEL STATUSREADY
A classifier should be evaluated on examples it did not use for training.
Responsible AI reflection
What could happen if the training images do not represent the real-world population?
Think about representation, lighting, camera quality, privacy and human review.