Supervised, Unsupervised, and Beyond

Where does the feedback come from?

Supervised learning consumes feature-label pairs and reduces a loss against those labels. Unsupervised learning has no answer sheet and searches for structure using assumptions about distance, density, or variation. The engineering question behind both is the same: what signal tells the system that it improved, and does that signal point toward the outcome we actually care about?

There is wide territory between supervised and unsupervised

Semi-supervised learning combines a small labeled set with a larger unlabeled set.
Self-supervised learning creates training targets from the input, such as next-token prediction or masked-patch recovery.
Reinforcement learning adjusts behavior from environmental reward, often with delayed credit.
Active learning chooses which uncertain examples a person should label next.

No labels does not mean no evaluation

If a human reviews and names clusters after every run, that judgment is part of the system. Sample original cases, ask whether members are actually similar, and determine whether the grouping changes a useful action. A high silhouette score proves only that one geometric criterion liked the partition.

The signal source predicts the failure mode

Human labels inherit human standards and bias. Distance-based structure changes with units and feature selection. A poorly designed reward can encourage the system to game the scoreboard instead of achieving the real goal. Record who creates the signal, when it arrives, how noisy it is, and how the method could optimize the wrong thing.

Usefulness appears in the next action

Test whether clustering improves support segmentation, whether a self-supervised representation improves a downstream classifier, or whether active learning reduces annotation cost. Internal scores show that a structure exists under the method's assumptions; downstream evidence shows whether that structure deserves a place in the product.

Code

Audit where the feedback signal lives·python

feedback_source = {
    "supervised": "explicit human labels in the dataset",
    "semi_supervised": "small labeled set + structure of unlabeled set",
    "self_supervised": "held-out part of the input itself",
    "unsupervised": "distance / density / structure assumptions",
    "reinforcement": "reward signal from environment",
    "active": "queries to a human oracle, ranked by uncertainty",
}

A simple semi-supervised setup with sklearn·python

from sklearn.semi_supervised import SelfTrainingClassifier
from sklearn.linear_model import LogisticRegression

base = LogisticRegression(max_iter=500)
ssl = SelfTrainingClassifier(base, threshold=0.85)
ssl.fit(X_partial, y_partial)  # y has -1 for unlabeled rows

Where does the feedback come from?

There is wide territory between supervised and unsupervised

No labels does not mean no evaluation

The signal source predicts the failure mode

Usefulness appears in the next action

Code

External links

Exercise

Progress

Comments 0