Domain Adaptation

Leverages a gradient reversal layer and a domain discriminator to learn feature representations that are indistinguishable between the source and target domains. The feature extractor is trained to maximize the loss of the domain classifier, effectively removing domain-specific information from the latent space.

• Core Mechanism: Minimax game between feature extractor and domain classifier
• Key Architecture: Domain-Adversarial Neural Network (DANN)
• Use Case: Aligning synthetic CAD-based defect images with real factory-floor camera feeds

A statistical approach that minimizes the distance between probability distributions in a Reproducing Kernel Hilbert Space (RKHS). MMD computes the squared distance between kernel embeddings of the source and target feature distributions, and this distance is added as a regularization term to the primary task loss.

• Core Mechanism: Non-parametric two-sample test used as a loss function
• Key Variants: Multi-kernel MMD (MK-MMD), Joint MMD (JMMD)
• Use Case: Transferring predictive maintenance models between different machine types with varying sensor configurations

Aligns the second-order statistics of source and target distributions by minimizing the Frobenius norm of the difference between their covariance matrices. CORAL applies a linear transformation to the source features to match the target covariance, requiring no labeled target data and no backpropagation through a domain classifier.

• Core Mechanism: Whitening and re-coloring of feature activations
• Key Advantage: Simple, efficient, and differentiable end-to-end
• Use Case: Adapting a quality inspection model to a new production line with different lighting conditions

An iterative, semi-supervised technique where a model initially trained on labeled source data generates pseudo-labels for unlabeled target data. High-confidence predictions are then added to the training set, and the model is retrained. This implicitly encourages the decision boundary to pass through low-density regions of the target feature space.

• Core Mechanism: Iterative bootstrapping using model's own predictions
• Key Risk: Confirmation bias if initial pseudo-labels are noisy
• Use Case: Bootstrapping a classifier on a new product SKU where only unlabeled real images are available

Finds a minimal-cost transportation plan to map source samples directly onto target samples based on a geometric cost function (e.g., Euclidean distance). Sinkhorn's algorithm is often used to compute an entropically regularized, differentiable approximation of the Wasserstein distance, enabling batch-wise alignment in deep learning pipelines.

• Core Mechanism: Probabilistic coupling between empirical distributions
• Key Variant: Joint Distribution Optimal Transport (JDOT)
• Use Case: Directly morphing synthetic depth maps to match the noise profile of a specific LiDAR sensor

Adapts a pre-trained model to a new target domain entirely at inference time, without accessing the source data. TTA typically updates batch normalization statistics on the target mini-batch or minimizes a self-supervised auxiliary task, such as rotation prediction or entropy minimization, to quickly recalibrate the model.

• Core Mechanism: Online, source-free model recalibration
• Key Constraint: Must operate with low latency and no target labels
• Use Case: A deployed edge model adjusting to gradual environmental drift, such as a factory's ambient temperature shift over a shift

The statistical divergence between the feature distributions of synthetic training data and real-world operational data. This gap is the fundamental problem that domain adaptation solves. It manifests as a drop in model accuracy when a system trained on pristine, rendered images encounters the noise, lighting, and texture variations of a physical factory floor. Quantifying this gap is the first step in any sim-to-real pipeline.

The end-to-end process of deploying a model trained entirely in simulation to a physical system. Domain adaptation is a critical sub-component of this pipeline. The workflow involves: 1) Building a digital twin of the target environment, 2) Generating synthetic data with domain randomization, 3) Applying domain adaptation to align feature distributions, and **4) Validating on a small real-world dataset. Success is measured by the sim-to-real transfer gap.

A metric for quantifying the quality and diversity of generated or adapted images. It compares the distribution of features extracted from a pre-trained Inception network for both the adapted synthetic images and real target images. A lower FID score indicates that the adapted images are statistically closer to the real distribution. It is the standard benchmark for evaluating domain adaptation quality.

A safety-critical technique for identifying inference-time inputs that differ fundamentally from the training distribution. After domain adaptation, a model may still encounter novel, unseen defect types or operational anomalies. OOD detection flags these inputs for human review rather than forcing a low-confidence prediction. It is the safety net that complements domain adaptation in production systems.