NeuHLS: A Neursoymbolic Framework for High-level Synthesis of Multi-Task Learning (NSF:SHF 2025-2029)
The growing demand for smart and autonomous systems has driven a surge in the deployment of edge devices. However, their limited computational resources and energy constraints often hinder the deployment of complex deep neural networks (DNNs). Optimizing DNNs for edge devices is crucial to unlock their full potential and enable a wider range of innovative applications. Our project aims to address the challenges of deploying DNNs on edge devices by developing a flexible and efficient framework. This project aims to exploit recent advances in multi-task learning, neurosymbolic AI, and high-level synthesis to develop new generation of tools that can automatically generate hardware accelerators for edge devices while satisfying latency and hardware platform constraints. In addition, the synthesized hardware must maximize the number of DNN weights implemented using software tunable parameters to allow for flexible fine-tuning at runtime. The proposed toolchain consists of three phases. In the first one, it aims to merge a set of single-task DNNs into one multi-task DNN by sharing representations from different single-task DNNs and hence reduces the model size. Next, the tool chain will exploit symbolic knowledge distillation (SKD) to compress the multi-task DNN into a neurosymbolic model which can be then processed by novel neurosymbolic high-level synthesis (HLS) techniques that will optimize the deployment while balancing accuracy, hardware utilization, and latency. The proposed toolchain will be rigorously evaluated using existing benchmarks and real-world application deployment in the domain of autonomous drones.