Low Power Autonomous and Smart Systems II – Invited Special Session

A5L-C: Low Power Autonomous and Smart Systems II - Invited Special Session

Session Type: Lecture
Session Code: A5L-C
Location: Room 3
Date & Time: Wednesday March 22, 2023 (15:20-16:20)
Chair: Tinoosh Mohsenin
Track: 12

Paper ID	Paper Name	Authors	Abstract
3216	Eventifying and Applying Bayesian Optimization in Lava	Maryam Parsa	Despite the monumental success of deep learning over the past few decades, researchers have yet to find a solution that overcomes the inherent fragility of machine learning models when facing adversarial attacks and stochastic environments. A variety of online, continual learning methods have been proposed to address these limitations with varying levels of success and cohesion. We introduce the first phase of the Eventified Bayesian Optimization (EBO) project implemented in Lava-Optimization, which aims to provide a generalized framework for eventified hyper-parameter optimization and lifelong continual learning applications to the neuromorphic research community. Here, we present an outline of the project over the next few years and highlight the completion of the first phase that implements a Bayesian optimization on Lava.
3217	Toward Metareasoning for Energy-Efficient Resource-Constrained Autonomous Systems	Mozhgan Navardi, Tinoosh Mohsenin	Safety, energy efficiency, low cost, and small size of smart drones have led to the proliferation of autonomous tiny UAVs in many indoor and outdoor applications. To make these tiny drones autonomous, Machine Learning (ML) algorithms show promising approaches in this area. However, deploying ML algorithms that demand high computational capacity on resource-constrained devices is challenging. For this aim, cloud-based approaches are highly considered although security concerns and communication challenges are still unaddressed. In this work, we implemented both cloud-based and edge computing approaches on a tiny drone. Moreover, we use a metareasoning approach as higher-level monitoring to improve the energy efficiency of autonomous tiny drones. We proposed an approach to address the communication and computation trade-off. In the proposed approach, the power consumption and latency regard to the communication and computation of cloud-based and the edge computing approaches are extracted. Therefore, the drone is able to switch between cloud-based implementation and edge computing to improve performance and energy efficiency by itself. A tiny drone called Crazyflie with an eight-core RISC-V processor is used for the experimental result.
3219	Multi-Objective Design Optimization for Image Classification Using Elastic Neural Networks	Lei Pan{2}, Yan Zhang{2}, Yi Zhou{1}, Heikki Huttunen{3}, Shuvra Bhattacharyya{2}	Image classification is an essential challenge for many types of autonomous and smart systems. With advances in Convolutional Neural Networks (CNNs), the accuracy of image classification systems has been dramatically improved. However, due to the escalating complexity of state-of-the-art CNN solutions, significant challenges arise in implementing real-time image classification applications on resource-constrained platforms. The framework of elastic neural networks has been proposed to address trade-offs between classification accuracy and real-time performance by leveraging intermediate early-exits placed in deep CNNs and allowing systems to switch among multiple candidate outputs, while switching off inference layers that are not used by the selected output. In this paper, we propose a novel approach for configuring early-exit points when converting a deep CNN into an elastic neural network. The proposed approach is designed to systematically optimize the quality and diversity of the alternative CNN operating points that are provided by the derived elastic networks. We demonstrate the utility of the proposed elastic neural network approach on the CIFAR-100 dataset.
3215	Efficient Multimodal Deep Neural Networks for Resource-Constrained Edge Devices	Hasib-Al Rashid, Tinoosh Mohsenin	With the rise of Artificial Intelligence (AI), there has been a renaissance in interest in how to apply AI algorithms on low-power embedded systems to expand possible Internet of Things (IoT) use-cases. Multimodal deep neural networks (M-DNN) have recently gained a lot of attention with the classification challenge due to their outstanding performance for computer vision and audio processing tasks. Their goal is to replicate multimodal human perception. This research focuses on a small, low-power software and hardware architecture for multimodal deep neural networks in edge devices with limited resources. Cyclic sparsification and hybrid quantization (4-bit weights and 8-bit activations) approaches are used to compress the models for implementation on small devices. We are the first to show the effectiveness of model compression strategies for multimodal deep neural networks employing cyclically sparsification and hybrid quantization of weights/activations, despite the fact that this is an active study topic. We provide three different case-studies running on various resource-constrained edge devices validating their energy-efficiency upon deployment.