Improving deep learning with prior knowledge and cognitive models: A survey on enhancing explainability, adversarial robustness and zero-shot learning

Incorporating domain knowledge into AI systems can address critical safety gaps: The paper identifies three major vulnerabilities in standard deep learning—adversarial attacks, lack of explainability, and poor generalization to new situations—and argues that adding explicit domain knowledge to these systems can help mitigate these problems in practical applications.

Current AI systems are vulnerable to adversarial attacks that exploit their data-driven nature: Deep learning models can be "tricked into making glaringly incorrect decisions" through adversarial attacks, a fundamental weakness that the paper suggests can be reduced by combining AI with domain expertise and cognitive models rather than relying solely on training data.

Brain-inspired approaches offer a path toward more interpretable AI: By mimicking how the human mind works, brain-inspired neural networks and cognitive architectures can enhance both the explainability of AI decisions and the intelligent behavior of autonomous systems, making AI systems more trustworthy and understandable to humans.

AI systems struggle with real-world generalization due to closed-world training assumptions: Standard deep learning models are typically trained on fixed datasets and fail when encountering genuinely new situations in open-world environments, a limitation the paper suggests can be addressed through zero-shot and few-shot learning approaches informed by prior knowledge.

Hybrid human-AI systems combining data and knowledge may be more reliable than purely data-driven approaches: The survey suggests that integrating domain knowledge, cognitive models, and brain-inspired methods with deep learning creates more robust, explainable, and generalizable AI systems for real-world deployment.