Does every training input need a label?

No. The model uses known labels where available and treats missing labels as latent for other examples.

Is an unlabeled example assigned one fixed class during training?

Not necessarily. The common variational objective averages over a predicted label distribution.

Can good image generation guarantee good classification?

No. The two goals are related through the model but require separate evaluation.

AI Glossary / Probabilistic machine learning / Intermediate

Semisupervised variational autoencoder

Definition

A semisupervised variational autoencoder is a latent-variable generative model trained using both labeled and unlabeled observations. In a common formulation, the class label is observed for some examples and treated as a latent variable for the rest. An inference network learns label probabilities while the generative model learns to explain the observations.

What it means

An organization may have many examples but only a small labeled subset. A semisupervised VAE attempts to use the unlabeled inputs' structure as well as the known labels. It does more than reconstruct data and attach a classifier afterward: in the common joint model, class information participates in generation and inference.

Several architectures fit the broad term. This entry focuses on the M2-style generative model introduced by Kingma and colleagues, where y is a class and z captures additional latent variation. The effectiveness of unlabeled data depends on whether the model's assumptions connect that variation with the classification task usefully.

How it works

The generative model specifies a class prior p(y), a latent prior p(z) and a decoder likelihood p(x|y,z). An inference model predicts q(y|x) and a distribution q(z|x,y). For a labeled example, the known y is used when forming the variational objective for the remaining latent variable.

For an unlabeled example, the objective averages class-conditioned bounds over q(y|x) and includes the entropy of that label distribution. With a manageable number of classes, the label expectation can be evaluated by summing over classes. Continuous latent variables can use a reparameterized gradient estimator.

The original approach also adds a supervised classification term so labeled examples train the classifier directly. The overall balance between labeled, unlabeled and classification contributions matters. At prediction time, q(y|x) supplies approximate class probabilities. These probabilities are learned model outputs, not guaranteed calibrated confidence or proof that the generative assumptions are correct.

Three practical examples

1. Document categorization: use a limited set of reviewed labels alongside a larger document collection. Check that the unlabeled collection uses the same relevant categories and does not overwhelm the model with unrelated material.

2. Image classification: model a class label and other visual variation jointly. Evaluate classification on held-out labels rather than judging the system only by reconstruction quality.

3. Sensor-event labeling: combine expert-labeled events with additional unlabeled sequences or feature vectors. Preserve a realistic evaluation split so repeated measurements from the same event do not leak across training and testing.

Think of it this way

Think of learning both how examples are constructed and which category descriptions fit them, using the few confirmed categories to orient the process. The analogy explains the combination of supervised and generative information. It does not imply that unlabeled examples reveal their true categories automatically or that visually coherent reconstructions establish correct classification.

Worked example

Consider one unlabeled input and two candidate classes A and B. Suppose their class-conditioned joint-data ELBO contributions are -4 and -6 nats. These values already include the class-prior term and the continuous-latent contribution. Let the inference classifier assign q(A|x)=0.75 and q(B|x)=0.25.

The weighted bound contribution is 0.75 × (-4) + 0.25 × (-6) = -4.5. The label entropy is -0.75 log(0.75)-0.25 log(0.25), approximately 0.562335. The unlabeled ELBO is therefore approximately -3.937665. Both classes contribute; the method has not simply assigned the example permanently to A.

If the same input had a known label A, its labeled generative bound would use the A contribution -4. An additional supervised classifier loss would be -log(0.75), approximately 0.287682, before applying its chosen weight. The example illustrates objective bookkeeping with stipulated values. It is not a complete training run, and the entropy term should not be mistaken for a claim that maximum label uncertainty is always desirable.

Illustrative case study

Fictional case: Birch Archive has a small reviewed set of document categories and a much larger unlabeled collection. Its team trains a semisupervised VAE and initially assumes that improved reconstruction means improved categorization. Reviewers discover that some unlabeled documents belong to categories absent from the labeled set.

The team investigates this mismatch, evaluates the classifier on independently reviewed examples and studies the weight of the unlabeled objective. It treats attractive generated samples as a separate result. This fictional case illustrates data and objective alignment, not evidence that adding unlabeled data necessarily improves a deployed classifier.

Common misconceptions and mistakes

Semisupervised does not mean self-supervised: known labels directly inform at least part of training. Nor is every VAE followed by a classifier the same joint architecture. In the original research, the latent-feature M1 approach and the generative M2 approach are distinct constructions.

Unlabeled data can hurt when its distribution or class coverage conflicts with the task assumptions. The additional latent variable z does not automatically contain only style, nor does y capture every meaningful category distinction. A good generative likelihood can reflect details irrelevant to classification. Class probabilities also need calibration checks when used for decisions or selective human review.

Business use

Consider the approach when labels are scarce, unlabeled inputs are relevant and modeling their structure is a plausible way to support classification. Compare against a supervised baseline and other semisupervised methods at the same labeling budget. Record how labels were selected, since a small convenience sample can miss important subgroups.

Separate reconstruction metrics, classification metrics and uncertainty calibration. Test class imbalance and distribution shift, and review whether new or unknown categories require explicit handling. The model's generative component adds assumptions and computational cost; it should earn its place through evaluation on the intended task.

Check your understanding

Why is the example's unlabeled ELBO not just the A bound of -4?

Answer: It combines both class-conditioned bounds according to q(y|x) and adds the label-distribution entropy. The label remains uncertain.

Related terms

Variational autoencoder
Reparameterization trick
semisupervised learning
latent class
evidence lower bound
inference network
label entropy
generative classifier

Learn next

Derive the labeled and unlabeled bounds for a two-class model, then trace the classifier's supervised loss. Next compare label-budget experiments and examine how distribution mismatch affects the usefulness of unlabeled examples.

Sources

Kingma and colleagues, Semi-Supervised Learning with Deep Generative Models, M2 construction and objectives: https://arxiv.org/pdf/1406.5298

The latent-label formulation and objective distinction were checked in the original paper. The numerical objective calculation and fictional archive case are original illustrations. Independent review consulted M2 equations 2, 4 and 6–9 for the labeled, unlabeled and classifier objectives.

Previous: Semisupervised parsing Next: Semivowel