Purely data-driven deep learning models are notoriously flexible, but this flexibility is a double-edged sword. When treated as unconstrained "black boxes," neural networks require massive datasets to learn basic physical laws, struggle to extrapolate beyond their training distributions, and frequently fail in low signal-to-noise environments. Conversely, classical engineering and scientific models—built on differential equations, physical constraints, and domain taxonomies—are robust and interpretable, but lack the capacity to adapt to complex, unmodeled real-world variations.
A rapidly growing paradigm bridges this gap by directly integrating domain knowledge into deep learning architectures. Rather than forcing a network to learn everything from scratch, this "model-driven" or "knowledge-informed" approach embeds established theoretical insights directly into the network's structure, loss functions, or data representations. By doing so, it constrains the model's search space, dramatically improving data efficiency, generalization, and robustness.
Constraining Neural Networks with Physical and Mechanistic Layers
Why does a generic neural network struggle with wireless communications or physical dynamics? Because the space of mathematically possible functions it can explore is infinitely larger than the space of physically possible functions. Knowledge-informed deep learning solves this by inserting non-trainable, mathematically explicit domain layers directly into the architecture.
Consider wireless image transmission over fading channels. Instead of letting a network attempt to learn the physics of signal degradation implicitly, Yang et al. in OFDM-Guided Deep Joint Source Channel Coding for Wireless Multipath Fading Channels and Deep Joint Source Channel Coding for Wireless Image Transmission with OFDM incorporate explicit Orthogonal Frequency Division Multiplexing (OFDM) and Minimum Mean Square Error (MMSE) channel estimation layers into a Joint Source-Channel Coding (JSCC) framework. Because the network does not waste capacity learning well-understood physical layer processing, it achieves superior image reconstruction under low signal-to-noise ratios (SNR). Similarly, Jiang et al. demonstrate in ISAC-NET: Model-Driven Deep Learning for Integrated Passive Sensing and Communication that embedding passive sensing and communication models into their network architecture yields enhanced robustness against communication errors.
This principle extends to ultra-reliable systems and dynamic physical processes. She et al. show in A Tutorial on Ultrareliable and Low-Latency Communications in 6G: Integrating Domain Knowledge Into Deep Learning that off-the-shelf deep learning cannot meet the stringent quality-of-service requirements of 6G networks without integrating communication theory into the simulation platforms and loss functions. When modeling continuous systems, Ravid et al. introduce GOKU-net in Generative ODE modeling with known unknowns, which embeds known ordinary differential equation (ODE) models into a generative framework. By restricting the network's latent trajectories to adhere to these mechanistic equations, the model can accurately infer "known-unknown" latent parameters and extrapolate far better than purely data-driven alternatives.
Unrolling Classical Optimization Algorithms
In signal and image processing, decades of research have produced highly effective iterative optimization algorithms. Rather than discarding these algorithms in favor of generic deep networks, researchers use a technique called algorithm unrolling.
As explained by Monga et al. in Algorithm Unrolling: Interpretable, Efficient Deep Learning for Signal and Image Processing, algorithm unrolling maps each iteration of a classical algorithm to a specific layer of a deep neural network. The structural priors and convergence insights of the original mathematical optimization are preserved within the network's architecture, while the specific parameters are learned from data. This hybrid approach yields models that are highly efficient, require significantly less training data, and remain fundamentally interpretable because their internal operations map directly to known optimization steps.
Injecting Common Sense via Structured Taxonomies and Knowledge Graphs
In complex scientific and clinical domains, data points are rarely independent; they exist within rich, hierarchical relationships. Purely sequential or pixel-based models often fail because they lack this "common sense." Integrating structured domain knowledge—such as medical taxonomies or chemical principles—provides a relational anchor.
In automated medical diagnosis, clinical text alone can be ambiguous. Yuan et al. address this in The Graph-based Mutual Attentive Network for Automatic Diagnosis by combining unstructured text with a Graph Convolutional Network that maps hierarchical medical knowledge and causal relationships between symptoms and diseases. This structured prior allows their model to capture complex clinical relationships that sequence-only models miss. For skin disease classification, Bajwa et al. demonstrate in Computer-Aided Diagnosis of Skin Diseases Using Deep Neural Networks that explicitly incorporating disease taxonomies into deep neural networks substantially improves diagnostic accuracy by helping the model differentiate between highly similar conditions.
To make these methods accessible, Yang et al. developed PyHealth: A Deep Learning Toolkit for Healthcare Applications, a toolkit designed to embed clinical taxonomies and medical knowledge graphs directly into deep learning pipelines. This integration has also proven vital for detecting complex psychological conditions like depression, where Pavlidis et al. highlight in Editorial: Special Issue on Unobtrusive Physiological Measurement Methods for Affective Applications the necessity of leveraging clinical knowledge graphs alongside raw physiological data.
Beyond healthcare, this structural injection works in chemistry. Generic language models often struggle to predict molecular structures from mass spectrometry data. Wohlwend et al. overcome this in Annotating metabolite mass spectra with domain-inspired chemical formula transformers by designing a chemical formula transformer that explicitly encodes chemical principles, such as neutral losses, outperforming generic architectures on complex scientific tasks.
Inductive Biases for Signal and Feature Refinement
Domain knowledge can also be integrated as targeted inductive biases within feature extraction and post-processing steps, acting as a powerful regularizer against noise.
For example, extracting heart rate from smartwatch photoplethysmogram (PPG) signals is notoriously difficult due to motion artifacts. Kechris et al. introduce KID-PPG in KID-PPG: Knowledge Informed Deep Learning for Extracting Heart Rate From a Smartwatch, which integrates expert signal processing and medical knowledge to adaptively remove motion noise and probabilistically assess signal degradation. In microscopy, Shen et al. demonstrate in DeepNoise: Signal and Noise Disentanglement Based on Classifying Fluorescent Microscopy Images via Deep Learning that incorporating experimental design priors into the post-processing of deep learning models allows for the precise disentanglement of biological signals from technical noise.
Similarly, in hyperspectral anomaly detection, Wang et al. utilize piecewise-smooth priors and saliency constraints to guide a graph autoencoder in Multi-Prior Graph Autoencoder with Ranking-Based Band Selection for Hyperspectral Anomaly Detection, ensuring the model focuses on physically meaningful spectral variations. For surgical tracking, Yao et al. combine Siamese networks with explicit geometric and imaging characteristics of guidewires to achieve robust real-time tracking in Real-Time Guidewire Tip Tracking Using a Siamese Network for Image-Guided Endovascular Procedures. Finally, in materials science, Yu et al. implement a feature-updating strategy based on the known physical nature of microstructural features to dramatically improve 3D atom probe tomography analysis in 3D deep learning for enhanced atom probe tomography analysis of nanoscale microstructures.
Why It Works
Knowledge-informed deep learning succeeds because it fundamentally re-aligns the learning problem. Instead of forcing a neural network to approximate a massive, unconstrained mapping from scratch, domain knowledge acts as a structural scaffolding. Whether through physical layers, unrolled algorithms, or hierarchical graphs, the architecture is pre-aligned with the underlying reality of the data. This leaves the network to do what it does best: learn the highly non-linear, data-driven residuals that classical models fail to capture.
Key Foundational Work
FingerNet: An unified deep network for fingerprint minutiae extraction (2017) — contributed 1 concept elements to this direction.
Deep Learning Techniques for Koala Activity Detection (2018) — contributed 1 concept elements to this direction.