[Paper Review] Knowledge-augmented Deep Learning and Its Applications: A Survey

Zijun Cui, Tian Gao+2 more|arXiv (Cornell University)|Nov 30, 2022

Domain Adaptation and Few-Shot Learning4 citations

TL;DR

This survey introduces knowledge-augmented deep learning (KADL) as a framework to enhance data efficiency, generalization, and interpretability in deep learning by integrating diverse domain knowledge—scientific and experiential—through systematic identification, representation, and integration. It presents a novel taxonomy of knowledge types and representations, offers a comprehensive review of integration methodologies across computer vision, NLP, and scientific modeling, and identifies key challenges and future directions, including hybrid integration, uncertainty modeling, and decision-level fusion.

ABSTRACT

Deep learning models, though having achieved great success in many different fields over the past years, are usually data hungry, fail to perform well on unseen samples, and lack of interpretability. Various prior knowledge often exists in the target domain and their use can alleviate the deficiencies with deep learning. To better mimic the behavior of human brains, different advanced methods have been proposed to identify domain knowledge and integrate it into deep models for data-efficient, generalizable, and interpretable deep learning, which we refer to as knowledge-augmented deep learning (KADL). In this survey, we define the concept of KADL, and introduce its three major tasks, i.e., knowledge identification, knowledge representation, and knowledge integration. Different from existing surveys that are focused on a specific type of knowledge, we provide a broad and complete taxonomy of domain knowledge and its representations. Based on our taxonomy, we provide a systematic review of existing techniques, different from existing works that survey integration approaches agnostic to taxonomy of knowledge. This survey subsumes existing works and offers a bird's-eye view of research in the general area of knowledge-augmented deep learning. The thorough and critical reviews of numerous papers help not only understand current progresses but also identify future directions for the research on knowledge-augmented deep learning.

Motivation & Objective

• To address the limitations of data-hungry, non-interpretable deep learning models by integrating domain knowledge.
• To define and systematize the three core tasks of KADL: knowledge identification, representation, and integration.
• To provide a broad, unified taxonomy of domain knowledge, including scientific and experiential knowledge, beyond narrow disciplinary surveys.
• To review integration techniques in a knowledge-type-aware manner, distinguishing between scientific and experiential knowledge applications.
• To identify open challenges such as uncertainty modeling, hybrid integration, and decision-level fusion for improved model robustness and interpretability.

Proposed method

• Proposes a novel taxonomy of domain knowledge, categorizing it into scientific knowledge (e.g., physical laws, mathematical equations) and experiential knowledge (e.g., rules, facts, heuristics).
• Classifies knowledge representation formats, including symbolic logic, knowledge graphs, probabilistic models, and mathematical equations.
• Systematically reviews knowledge integration techniques, distinguishing between data-driven training-time integration and decision-level fusion using prior models.
• Categorizes integration methods by knowledge type, such as physics-informed neural networks for scientific knowledge and neural-symbolic models for experiential knowledge.
• Introduces a prescriptive tree (Figure 9) mapping knowledge types to representation formats and integration methods, serving as a roadmap for KADL design.
• Emphasizes the need for uncertainty-aware integration, advocating for probabilistic graphical models to capture uncertainty in both scientific and experiential knowledge.

Experimental results

Research questions

• RQ1 How can diverse types of domain knowledge—scientific and experiential—be systematically identified and categorized in deep learning applications?
• RQ2 What are the most effective representation formats for different types of domain knowledge, and how do they influence model performance?
• RQ3 How can knowledge be integrated into deep learning models in a way that enhances data efficiency, generalization, and interpretability?
• RQ4 What are the limitations of current integration methods, particularly in handling uncertainty and combining multiple knowledge types?
• RQ5 What future research directions are most promising for achieving robust, generalizable, and interpretable knowledge-augmented deep learning?

Key findings

• The survey establishes a comprehensive taxonomy of domain knowledge, distinguishing scientific knowledge (e.g., physical laws) from experiential knowledge (e.g., rules, heuristics), which subsumes prior narrow-scope surveys.
• Existing integration methods are predominantly trained jointly with data, lacking principled separation between data and knowledge, which limits interpretability and flexibility.
• Decision-level fusion—where a prior model based on knowledge combines with data-driven predictions—offers a more modular and interpretable alternative to end-to-end training-time integration.
• Hybrid integration methods that combine multiple techniques (e.g., architectural design, regularization, and refinement) are underexplored but promising for leveraging complementary strengths.
• Uncertainty in both scientific and experiential knowledge is largely ignored in current methods, and probabilistic modeling of such uncertainty remains an open challenge.
• The survey identifies a critical gap in real-world application-focused KADL research, as most existing works are limited to lab-scale scientific problems rather than broad, practical deployment.