Real-time and efficient neurosymbolic AI on edge devices

Establish methods that enable real-time and efficient execution of neurosymbolic artificial intelligence workloads on resource-constrained edge devices, ensuring end-to-end neurosymbolic inference meets real-time performance and efficiency requirements.

Background

The paper motivates neurosymbolic AI as a compositional paradigm that integrates neural perception with symbolic reasoning, often implemented using vector-symbolic architectures (VSA). While such systems achieve strong reasoning performance, they are difficult to deploy efficiently on existing hardware due to large memory footprints (e.g., symbolic codebooks), heterogeneous compute kernels (e.g., circular convolutions), and irregular, sequential processing that underutilizes accelerators like GPUs/TPUs.

The authors highlight that, despite advances demonstrated on powerful server-class hardware, achieving real-time and efficient neurosymbolic inference on edge devices remains challenging under tight resource constraints. They position this as a central open problem motivating their algorithm-hardware co-design approach (CogSys).