This paper reviews the state-of-the-art (SOTA) deep learning (DL) models for IoT and identifies their limitations, particularly in terms of testability, verifiability, and interpretability. It explores how neuro-symbolic methods can address these challenges. The paper discusses key challenges and research opportunities in enhancing AIoT reliability with neuro-symbolic approaches, including hard-coded symbolic AI, multimodal sensor data, biased interpretability, trade-offs between interpretability and performance, and ethical and societal issues. The neuro-symbolic paradigm combines the robustness of symbolic AI with the flexibility of DL, enabling AI systems to reason, make decisions, and generalize knowledge from large datasets. The paper also identifies the remaining neuro-symbolic challenges and discusses the potential of neuro-symbolic approaches in addressing the limitations of DL in AIoT. The paper reviews neuro-symbolic approaches and their applications, focusing on their potential to address the challenges faced by DL models in AIoT. The main contributions of the paper are: identification of SOTA DL models used in AIoT applications and an in-depth exploration of their limitations, examination of how neuro-symbolic methods can offer solutions to the issues in DL models, and identification of key challenges and research opportunities in enhancing AIoT reliability with neuro-symbolic approaches. The paper is organized as follows: Section 2 overviews SOTA DL models used in IoT, the challenges of which are outlined in Sect. 3. Section 4 introduces the neuro-symbolic paradigm, including the evolutionary stages of neural and symbolic AI, SOTA approaches, applications, and their key advantages. Finally, Sect. 5 describes the remaining challenges of neuro-symbolic techniques in AIoT. The paper discusses various DL models used in IoT, including CNNs, RNNs and LSTMs, RBMs, AEs, GNNs, geometric deep learning models, and transformer models. It also discusses the challenges of DL models, including testability, interpretability, and verifiability. The paper highlights the need for neuro-symbolic approaches to address these challenges and improve the reliability of AIoT systems. The paper concludes that neuro-symbolic approaches offer a promising solution for creating more robust and interpretable AI systems.This paper reviews the state-of-the-art (SOTA) deep learning (DL) models for IoT and identifies their limitations, particularly in terms of testability, verifiability, and interpretability. It explores how neuro-symbolic methods can address these challenges. The paper discusses key challenges and research opportunities in enhancing AIoT reliability with neuro-symbolic approaches, including hard-coded symbolic AI, multimodal sensor data, biased interpretability, trade-offs between interpretability and performance, and ethical and societal issues. The neuro-symbolic paradigm combines the robustness of symbolic AI with the flexibility of DL, enabling AI systems to reason, make decisions, and generalize knowledge from large datasets. The paper also identifies the remaining neuro-symbolic challenges and discusses the potential of neuro-symbolic approaches in addressing the limitations of DL in AIoT. The paper reviews neuro-symbolic approaches and their applications, focusing on their potential to address the challenges faced by DL models in AIoT. The main contributions of the paper are: identification of SOTA DL models used in AIoT applications and an in-depth exploration of their limitations, examination of how neuro-symbolic methods can offer solutions to the issues in DL models, and identification of key challenges and research opportunities in enhancing AIoT reliability with neuro-symbolic approaches. The paper is organized as follows: Section 2 overviews SOTA DL models used in IoT, the challenges of which are outlined in Sect. 3. Section 4 introduces the neuro-symbolic paradigm, including the evolutionary stages of neural and symbolic AI, SOTA approaches, applications, and their key advantages. Finally, Sect. 5 describes the remaining challenges of neuro-symbolic techniques in AIoT. The paper discusses various DL models used in IoT, including CNNs, RNNs and LSTMs, RBMs, AEs, GNNs, geometric deep learning models, and transformer models. It also discusses the challenges of DL models, including testability, interpretability, and verifiability. The paper highlights the need for neuro-symbolic approaches to address these challenges and improve the reliability of AIoT systems. The paper concludes that neuro-symbolic approaches offer a promising solution for creating more robust and interpretable AI systems.