This survey comprehensively reviews the recent advances in the area of Neurosymbolic Artificial Intelligence (AI), a hybrid approach that integrates the learning capacity of neural networks and the reasoning capability of symbolic AI. Fourteen major papers have been systematically analyzed, emphasizing systems constructed with prior knowledge, significantly reducing the data requirements for achieving satisfactory performance concerning traditional metrics like accuracy.

The survey presents a detailed taxonomy of different aspects within Neurosymbolic AI, including Knowledge Representation, Learning and Reasoning. An overview of the types of queries Neurosymbolic Question-Answering (NSQA) systems can effectively answer is also provided. This includes simple questions, multi-relational inquiries, count-based, superlative, comparative, geographic, and temporal questions.

Despite the benefits of Neurosymbolic AI, certain limitations were identified. These include the inability to save models in Logic Tensor Networks (LTNs) without re-training, a lack of user-friendly explainability, and an absence of robust comparative studies between neurosymbolic and non-logic/symbolic approaches, such as deep learning and decision trees. Furthermore, the study highlighted the under-explored potential of Neurosymbolic AI in addressing fairness and algorithmic bias.

While Neurosymbolic AI systems often underperform compared to pure Deep Learning techniques, the benefits such as reduced data requirements, computational cost efficiency, and enhanced comprehensibility warrant further exploration as a potential alternative to ever-expanding deep learning models.