Why it matters
The paradigm determines the data requirement, the evaluation method and the failure mode. Supervised models fail when labels are noisy or unrepresentative; self-supervised models fail when the pretext task does not require the capability you need; reinforcement learning fails when the reward is misspecified or exploration is expensive. Choosing a paradigm is therefore a statement about what signal you can actually obtain at scale.
Mental model
The paradigms are different ways to learn a craft: apprenticeship with a master (supervised), studying the material without answers (unsupervised), practising by predicting missing steps (self-supervised), and trial with reward (reinforcement). The metaphor breaks because the same model can pass through several of these stages, so the paradigm is a phase of training rather than a type of model.
How it works
- 1Ask what supervision is available at the scale you need: labels, raw structure, or a simulator with reward.
- 2For supervised work, estimate label cost and noise before choosing model capacity.
- 3For self-supervised work, define the pretext task and check that solving it requires the representation you want.
- 4For unsupervised work, choose a downstream proxy metric, since the training objective does not measure usefulness.
- 5For reinforcement learning, verify that the reward is measurable, dense enough and hard to game.
- 6Combine stages when appropriate, such as pretraining self-supervised and adapting with a small labelled set.
Step 1 of 6
Ask what supervision is available at the scale you need: labels, raw structure, or a simulator with reward.
Figure. The mechanism in this page's own steps. The figure adds sequence, not new claims.
What to notice. A step can be read alone, but the order is the mechanism. The control lets you stop on any one.
Text version
The page's steps: 1. Ask what supervision is available at the scale you need: labels, raw structure, or a simu… 2. For supervised work, estimate label cost and noise before choosing model capacity. 3. For self-supervised work, define the pretext task and check that solving it requires the … 4. For unsupervised work, choose a downstream proxy metric, since the training objective doe… 5. For reinforcement learning, verify that the reward is measurable, dense enough and hard t… 6. Combine stages when appropriate, such as pretraining self-supervised and adapting with a …
- Ask what supervision is available at the scale you need: labels, raw structure, or a simulator with reward.
- For supervised work, estimate label cost and noise before choosing model capacity.
- For self-supervised work, define the pretext task and check that solving it requires the representation you want.
- For unsupervised work, choose a downstream proxy metric, since the training objective does not measure usefulness.
- For reinforcement learning, verify that the reward is measurable, dense enough and hard to game.
- Combine stages when appropriate, such as pretraining self-supervised and adapting with a small labelled set.
Trade-offs
| Choice | Good for | Cost | Fails when |
|---|---|---|---|
| Supervised | Direct optimisation of the target metric | Label cost, brittle to distribution shift | Labels are scarce or noisy |
| Self-supervised | Scales with unlabelled data, strong transfer | Pretext task may not match the downstream need | You need calibrated outputs for a narrow task |
| Unsupervised | No labels, discovers structure | Evaluation is indirect and ambiguous | The product needs a specific prediction |
| Reinforcement | Optimises long-horizon reward directly | Sample inefficiency, reward hacking | The environment is expensive or unsafe |
Good for
Direct optimisation of the target metric
Cost
Label cost, brittle to distribution shift
Fails when
Labels are scarce or noisy
- Labels are scarce or noisy
- You need calibrated outputs for a narrow task
- The product needs a specific prediction
- The environment is expensive or unsafe
Figure. Each choice with what it buys, what it costs, and the case where it loses.
What to notice. The fourth column is the decision: if the losing case describes your situation, the choice is wrong.
Text version
Four trade-offs: Supervised is good for Direct optimisation of the target metric, costs Label cost, brittle to distribution shift, and fails when Labels are scarce or noisy. Self-supervised is good for Scales with unlabelled data, strong transfer, costs Pretext task may not match the downstream need, and fails when You need calibrated outputs for a narrow task. Unsupervised is good for No labels, discovers structure, costs Evaluation is indirect and ambiguous, and fails when The product needs a specific prediction. Reinforcement is good for Optimises long-horizon reward directly, costs Sample inefficiency, reward hacking, and fails when The environment is expensive or unsafe
Failure modes
- Pretext task too easy. The model solves the surrogate without learning useful structure, and transfer is weak. Debug: probe the learned representation on a small labelled set and compare with a supervised baseline.
- Label noise treated as truth. Supervised training fits annotator mistakes and validation agrees because it uses the same annotators. Debug: measure inter-annotator agreement and inspect disagreements.
- Reward misspecified. The policy optimises the measured proxy and degrades the real objective. Debug: evaluate on a held-out human judgement not used in the reward.
In production
State which paradigm produced each checkpoint, because evaluation and retraining depend on it. For self-supervised and unsupervised models, keep a small labelled probe set so you can detect representation drift without a full downstream retrain.
Common misconceptions
People think: Unsupervised learning needs no evaluation.
Actually: Without a target the training objective cannot tell you whether the structure is useful, so a downstream proxy is still required.
People think: Reinforcement learning is the general solution when labels are missing.
Actually: It needs a reliable reward and many interactions; when a reward model can be trained from comparisons, supervised preference learning is often cheaper.
Interview angle
A strong answer maps each paradigm to its supervision source, cost and failure mode, and explains why self-supervised pretraining followed by supervised adaptation is common. It should discuss evaluation for unsupervised work. Weak answers define the paradigms by whether labels exist without saying what signal replaces them or how you would know the model learned something useful.
Sources
- Sutton & Barto, Reinforcement Learning: An Introduction — the reward-based learning framework.
- Bengio et al., Representation Learning: A Review and New Perspectives — self-supervised and unsupervised representation learning.
- Murphy, Probabilistic Machine Learning: An Introduction — supervised, unsupervised and reinforcement formulations.