Asynchronous Neuromorphic Optimization with Lava

TL;DR

The paper tackles the challenge of performing optimization on event-based neuromorphic systems where cross-device asynchrony can cause deadlocks. It introduces an asynchronous Bayesian optimization framework within the Lava ecosystem, designed to be Loihi 2-compatible and scalable to other optimization methods. A key contribution is an intermediate handshake mechanism that coordinates between the optimizer and a black-box function, enabling non-blocking communication via Stop/Pause commands, completion signaling, and sleep-probe behavior. Demonstrated on a graph-based satellite scheduling problem solved with a QUBO model on Loihi 2, the approach shows robust, asynchronous interaction across CPU and neuromorphic hardware, reducing wasted cycles and enabling real-time-like operation. The work lays the groundwork for multi-agent asynchronous optimization and lifelong on-chip learning in robotics and signal processing, backed by Intel support.

Abstract

Performing optimization with event-based asynchronous neuromorphic systems presents significant challenges. Intel's neuromorphic computing framework, Lava, offers an abstract application programming interface designed for constructing event-based computational graphs. In this study, we introduce a novel framework tailored for asynchronous Bayesian optimization that is also compatible with Loihi 2. We showcase the capability of our asynchronous optimization framework by connecting it with a graph-based satellite scheduling problem running on physical Loihi 2 hardware.

Figures (2)

• Figure 1: State diagrams comparing optimizer to black-box function communication with synchronous and asynchronous operation. (A) Synchronous communication where the optimizer transmits a single hyperparameter value $X'$. The black-box function receives $X'$ and calculates $Y$ via $Y=F(X')$ before returning the tuple $(X,Y)$. (B) The same scenario as A but where the black-box function takes multiple steps to calculate the result $Y$. This causes a programmatic deadlock where the optimizer is receiving data that isn't available. (C) Using a probe, the optimizer will determine if the result is fully processed from the black-box function. If the data isn't available, the process will proceed to the next timestep.
• Figure 2: The communication loop between our asynchronous optimization framework in Lava and the satellite scheduling problem. $X'$ is an unknown set of parameters from the search algorithm where the Satellite Scheduling Runtime evaluates $X'$ and returns the number of scheduled satellites $Y$. The resulting tuple $(X, Y)$ will arrive asynchronously with respect to the configuration, compilation, scheduling, and execution of the satellite scheduler on Loihi.