CONIX Publication

Neuro-symbolic Architectures for the Internet of Things

Authors: Mani Srivastava


The nexus of deep neural networks (DNNs) and the Internet of Things (IoT) allows sensing and actuation to be performed in our personal, social, and physical spaces in previously unimagined ways. Deep learning methods deployed across the edge-cloud continuum enable IoT systems to make accurate predictions and decisions from high-dimensional and unstructured real-world sensory data while benefiting from the high-performance tensor operations in hardware accelerators. As a result, in many settings, DNNs have entirely replaced symbolic and mechanistic approaches based on algorithms, scientific models, and human knowledge. However, the benefits come with considerably reduced abilities to generalize to new situations, to assure trustworthiness, and to reason about complex spatiotemporal events that require connecting the dots across large spans of time and space. We will present emerging neuro-symbolic approaches that seek to overcome this tension by integrating neural representations with symbolic reasoning. The former allows efficient processing of multimodal sensory inputs to create precepts that assist reasoning and the latter provides interpretability, enforces constraints, allows for human knowledge injection, and acts as regularizers that guide the learning of neural components. The talk will describe the unique capabilities that neuro-symbolic architectures bring to the IoT domain and the research challenges they present.

Release Date: 06/27/2022
Uploaded File: View