How Can We Trust an Unsupervised Model?
Unsupervised models are evaluated by measurable properties of the structure they produce — internal consistency, stability under perturbation, and downstream usefulness — rather than by matching a hidden ground truth.
A complete interactive classroom, not just a preview.
Start when you are ready to enter this Stage's 7 scenes and explore, respond, and learn as you go.
How can we tell that an unsupervised model is correct when there is no labeled 'right answer' to compare against?
A model can group your customers, segment your cells, or compress your images — all without you ever telling it the 'right' answer. So how would you know it got it right?
Supervised models are judged by ground-truth labels. Unsupervised models invent their own structure. The intuition that 'no labels means no way to verify' feels obvious — and is misleading.
A hands-on clustering simulator where the learner reshapes data and watches internal validation metrics rise and fall, revealing that 'rightness' in unsupervised learning is a measurable property of the structure itself, not a hidden label.
Unsupervised models are checked by asking whether their output structure is stable, coherent, and informative — internal evidence replaces external labels.
Without labels, any output is just a guess; there's no principled way to say one clustering or embedding is better than another.
- Full taxonomy of clustering algorithms (k-means internals, DBSCAN details, hierarchical linkage)
- Deep unsupervised methods in depth (autoencoders, contrastive objectives, GAN training)
- Statistical hypothesis testing for arbitrary cluster counts
- Semi-supervised and self-supervised learning as separate paradigms
- 01The Verification GapslideQuestion
Pose the driving question: if no one tells the model what the answer is, how can the model — or we — know it produced something meaningful?
- Supervised models win or lose against labeled ground truth
- Unsupervised models invent structure without labels
- The verification problem is real but not unsolvable
- 02Your First InstinctquizPrediction
Before any evidence, commit to one answer to the driving question.
- Choose the strategy you think actually works in practice
- Lock in your initial intuition so the evidence scene can confront it
- 03Watch Clusters Earn a ScoreinteractiveEvidence
Reshape a 2D point cloud — spread clusters apart, merge them, or add noise — and watch an internal validation score change in real time.
- Cluster separation changes the score without any labels being introduced
- Noise and overlap lower the score even though no 'answer' is given
- The structure itself produces measurable evidence
- 04Three Kinds of Internal EvidenceslideExplanation
Explain why internal metrics, stability, and downstream usefulness together replace the role of ground-truth labels.
- Internal metrics: cohesion within groups and separation between groups
- Stability: the same structure reappears under resampling and perturbation
- Downstream utility: the structure makes a later supervised task easier
- 05Test for StabilityinteractiveTransfer
Resample the same dataset and watch how a clustering assignment shifts. Learners judge whether the structure is robust or fragile.
- Compare two runs of the same algorithm on resampled data
- A robust structure shows high agreement between runs
- Fragility is evidence the model is fitting noise, not signal
- 06Where Internal Evidence LiesslideBoundary
Show the limits: internal metrics can reward obviously wrong structure when the data has no real clusters, and stability can hide trivial solutions.
- A single tight cluster can score perfectly on cohesion even when the data is one blob
- Stability alone cannot distinguish 'meaningful' from 'boringly constant'
- Evaluation always needs more than one signal
- 07So How Do We Know?slideResolution
Directly answer the driving question by reuniting the three evidence types into a single verification posture.
- Unsupervised models are checked by the structure they produce, not by hidden labels
- Combine internal metrics, stability, and downstream utility
- Confidence in an unsupervised model comes from converging evidence, not from a single number
Discussion threads for a Stage aren't available yet.
