Domain Adaptation

Directly minimizes statistical divergence between source and target feature distributions using explicit distance metrics.

• Maximum Mean Discrepancy (MMD) measures the distance between kernel embeddings of source and target distributions in a reproducing kernel Hilbert space
• Correlation Alignment (CORAL) aligns second-order statistics by whitening and re-coloring feature covariances
• Central Moment Discrepancy (CMD) matches higher-order moments to capture complex distributional shifts

Applied when a factory's lighting conditions or camera angles differ systematically from the training dataset, requiring explicit statistical realignment.

Employs a domain discriminator network trained adversarially against the feature extractor to produce domain-invariant representations.

• A gradient reversal layer (GRL) flips gradients during backpropagation, forcing the feature extractor to confuse the domain classifier
• The feature extractor learns to strip away domain-specific information while preserving task-relevant semantics
• Domain-adversarial neural networks (DANN) represent the canonical architecture for this approach

Effective when the target domain is unlabeled, such as adapting a defect classifier to a new production line without annotated examples.

Iteratively generates pseudo-labels for unlabeled target data using the model's own confident predictions, then retrains on these augmented examples.

• Only predictions exceeding a confidence threshold are retained to prevent error propagation
• Curriculum learning strategies gradually increase the proportion of pseudo-labeled data as model confidence improves
• Consistency regularization enforces that the model produces identical predictions for perturbed versions of the same input

Particularly valuable for rare defect detection where labeling target-domain anomalies is prohibitively expensive.

Deliberately varies non-essential parameters of the training environment—such as lighting intensity, camera position, texture, and background—to force the model to learn invariant features.

• The model encounters such extreme diversity during training that the real target domain appears as just another variation
• Originated in sim-to-real transfer for robotics but now widely applied to visual inspection systems
• Requires careful parameterization of the randomization space to avoid training on unrealistic configurations

Enables a single model to deploy across multiple factory sites with different physical setups without per-site fine-tuning.

Adapts model parameters at inference time using only the unlabeled target sample or a small batch, without access to source data.

• Batch normalization adaptation updates running statistics (mean, variance) on target data while freezing all other weights
• Entropy minimization encourages confident predictions by reducing output distribution entropy on target samples
• TENT (Test Entropy) modulates only batch normalization affine parameters via entropy minimization

Critical for dynamic manufacturing environments where conditions shift continuously—such as gradual sensor drift or changing ambient temperature throughout a production shift.

Leverages a small number of labeled target examples (typically 1-10 per class) to rapidly align the model to the new domain.

• Prototypical networks compute class prototypes in embedding space from the few labeled examples and classify based on distance to these prototypes
• Matching networks use attention mechanisms over the small support set to produce predictions for query samples
• Combines with fine-tuning strategies like LoRA to update only a minimal parameter subset

Ideal when a process engineer can quickly annotate a handful of examples from the new production environment to bootstrap adaptation.