Distributed actor-critic method for multi-objective resource-aware neural architecture search

In the era of diverse hardware platforms, a critical element of neural network architecture optimization is ensuring that the architecture is optimized for a specific deployment environment. This paper aims to present an effective distributed method for optimizing neural architectures to address the challenge of achieving multiple objectives simultaneously. Using the power of distributed computing, our method employs an Actor-Critic framework to explore the architecture search space while considering multiple resource-related objectives such as memory footprint for deployments in memory-constrained systems, and the amount of computation required given a number of Floating-point Operations Per Second (FLOPs) for applications deployed on energy-constrained devices. An important innovation that is proposed in this research is the integration of hardware awareness, multi-objective optimization, and multi-modality in data. It enables neural architectures to meet strict hardware constraints, including memory efficiency and inference speed, while handling a wide range of data modalities. The proposed method was evaluated in four scenarios, incorporating 3 image, 4 text, 8 graph, and 8 time series benchmarks. Despite the diversity of data modalities, the method demonstrates consistent performance and adaptability.