Presentation
MERINDA: Model Recovery in FPGA based Dynamic Architecture
DescriptionRecovering underlying governing equations, i.e., model recovery (MR) from data - crucial for runtime monitoring - is one of the key solutions for assurance of safe and explainable operations of mission-critical autonomous systems (MCAS). MCAS often operate under strict constraints related to time, computational resources, and power, potentially requiring the usage of edge artificial intelligence (edge-AI) for accelerated safety monitoring. Field Programmable Gate Arrays (FPGAs) have emerged as an ideal solution to meet these constraints due to their reconfigurability and capacity for hardware-optimized performance. MR approaches such as EMILY or Physics informed neural networks with sparse regression (PINN+SR), use continuous depth residual networks and continuous time latent variable models e.g. Neural ODE (NODE) as core components. A key challenge in accelerating these components comes from the iterative approach towards integration of ordinary differential equations (ODE) in the forward pass. This paper introduces a novel FPGA-based accelerated model recovery in dynamic architecture (MERINDA) approach, that utilizes equivalent neural architectures to the NODE core that are amenable for acceleration. MERINDA has four components: a) a Gated Recurrent Unit (GRU) layer that generates approximate discretized solution of the original NODE layer in EMILY or PINN-SR, b) a dense layer to solve the inverse ODE problem to obtain approximate ODE model coefficients from the discretized solutions, c) sparsity driven dropout layer to reduce model order, and d) a standard ODE solver to solve the reduced order ODE and regenerate the continuous time original output of the NODE layer. Components a and b can be accelerated using FPGA, while c and d have much less computational complexity than the original NODE layer. We first theoretically prove that the MERINDA is equivalent to EMILY and empirically compare their MR performance on four benchmark examples. We then assess the accuracy, processing speed, energy consumption, and DRAM access of MERINDA and compare with baseline accelerated machine learning approaches such as non-physics informed machine learning (ML), and learning with only physics guided loss functions (ML-PG). We also compare MERINDA with GPU-based implementations of all three approaches to evaluate the advantage of acceleration. Finally, we apply mixed integer programming to identify the optimal approach towards runtime monitoring and its hyper-parameters under various resource constraints. Our results demonstrate substantial improvements in energy efficiency, training time, and DRAM access with minimal compromises in accuracy, underscoring the viability of MERINDA for resource-constrained autonomous systems.
Event Type
Networking
Work-in-Progress Poster
TimeSunday, June 226:00pm - 7:00pm PDT
LocationLevel 3 Lobby