Running multiple deep neural networks (DNNs) simultaneously on mobile devices introduces challenges due to constrained computing resources. Previous research has explored the use of heterogeneous processors for accelerating DNN inference but often overlooks thermal issues, which can degrade computing power. In this article, we propose Phoenix, a system specifically designed to enhance the performance of multi-instance DNNs in video applications by maximizing accuracy and ensuring the achievement of a required frame rate. Phoenix allocates DNN tasks to the most suitable hardware processors, understanding complex thermal dynamics through reinforcement learning, and postpones the onset of thermal throttling. Despite optimized task allocation, continuous inference of multiple DNNs can still lead to thermal throttling. To manage performance degradation, Phoenix employs a multi-exit network, adaptively executing inference tasks to ensure consistent frame rates. Phoenix minimizes accuracy loss from early exits by optimally generating and operating multi-exit networks. We evaluated Phoenix using two different benchmarks and Virtual Youtuber streaming application. The results demonstrated that Phoenix effectively enhances device performance by delaying thermal throttling and achieving optimal accuracy while maintaining a consistent frame rate.

Deep neural networks (DNNs) have enabled a wide range of artificial intelligence of things (AIoT) applications, but their increasing complexity poses challenges for deployment on resource-constrained devices. Model compression techniques such as pruning and quantization have been widely adopted to address these challenges; however, they inevitably incur accuracy loss due to information loss. Recently, class-aware pruning has emerged as a promising approach, but existing methods often lack flexibility, as they are typically tailored to fixed target class sets and fail to generalize well to dynamic or broad class distributions. To address this limitation, we propose Mixture of Class-aware Experts (MoCE), a novel framework that combines class-aware pruning with a Mixture of Experts (MoE) architecture. MoCE constructs multiple lightweight experts using class-aware pruning, each specialized for a subset of classes, and employs a shared encoder and a lightweight router to dynamically select the appropriate expert at runtime. Our preliminary results demonstrate the potential of combining class-aware pruning and expert selection to enable accurate and efficient inference on resource-limited AIoT devices.

Optimizing deep neural networks (DNNs) running on resource-constrained devices, such as energy harvesting sensor devices, poses unique challenges due to the limited memory and varying energy conditions. Existing efforts have shown that deploying a multi-exit network mitigates the problem by allowing tradeoffs between accuracy and computational complexity. However, previous works did not fully consider two essential requirements: optimized neural architecture and optimized inference policy. In this paper, we present HarvNet, which comprises two complementary techniques for generating and operating a multi-exit network for energy harvesting devices. First, we provide a neural architecture search scheme, HarvNAS, which configures the best multi-exit architecture while meeting memory and energy constraints. Second, HarvSched learns and constructs the best progressive inference policy with different energy constraints by considering runtime factors, such as the harvesting status and the energy storage level. We implemented HarvNAS using the TensorFlow framework and then implemented and evaluated HarvSched on an MSP430-based sensor device. The evaluation showed that HarvNAS generated a model with up to 2.6%p higher accuracy while saving up to 70% of memory compared to the existing technique, and HarvSched enabled zero-downtime operation of the generated model.

Drones are commonly used in mission-critical applications, and the accurate estimation of available battery capacity before flight is critical for reliable and efficient mission planning. To this end, the battery voltage should be predicted accurately prior to launching a drone. However, in drone applications, a rise in the battery's internal temperature changes the voltage significantly and leads to challenges in voltage prediction. In this paper, we propose a battery voltage prediction method that takes into account the battery's internal temperature to accurately estimate the available capacity of the drone battery. To this end, we devise a temporal temperature factor (TTF) metric that is calculated by accumulating time series data about the battery's discharge history. We employ a machine learning-based prediction model, reflecting the TTF metric, to achieve high prediction accuracy and low complexity. We validated the accuracy and complexity of our model with extensive evaluation. The results show that the proposed model is accurate with less than 1.5% error and readily operates on resource-constrained embedded devices.

Battery-powered devices commonly use Li-ion batteries due to their high energy density. Unfortunately, low ambient temperature and high discharge current significantly affect the available capacity of Li-ion batteries at runtime. One solution to handle this issue is to construct a hybrid energy storage (HES) system, taking advantage of the different performance characteristics of various Li-ion batteries. Conventional HES techniques often configure systems without considering the hardware overhead and capacity loss of batteries. Moreover, dynamic energy management in HES was not considered adequately due to severe power leakage. In this article, we propose DynLiB, a reconfigurable HES architecture that provides high energy availability in various runtime environments. DynLiB constructs HES with only two types of batteries, the main battery and the auxiliary battery, to reduce the operational cost of the hardware. The capacity of each battery is determined based on the application profile that characterizes the runtime factors. DynLiB supports energy transfer between batteries so that the available energy is maximized by dynamically transferring energy between the batteries, depending on the runtime environment. We implemented a prototype system to validate the efficacy of DynLiB. The experimental results showed that DynLiB achieved an average energy gain of 27.2% compared to a single-type battery system in various runtime environments.

Multiple photovoltaic (PV) modules are often used to provide enhanced harvesting capability for light energy-based Internet of Things (IoT) devices. PV modules facing multiple directions can lead to a situational energy loss when parts of the modules are shaded. To address this issue, existing solutions exploit reconfigurable PV arrays to acquire the optimal configuration in a given situation. However, conventional techniques are not energy efficient in estimating the harvesting capability of PV modules, and require high computation to find the optimal PV array at runtime. In this article, we propose PVoT, an energy-efficient reconfigurable PV array, which maximizes the harvesting energy for indoor IoT devices. To this end, we propose the use of photoresistors to estimate the harvesting capability with minimal energy overhead. We also provide hardware and software schemes, which perform event-driven light change detection in an energy-efficient way. Furthermore, we develop a power imbalance threshold metric to quickly find the optimal PV array at runtime. We implemented a prototype PVoT with off-the-shelf components and accompanying software. Experiments with the prototype hardware showed that PVoT achieves a gain of up to 23.9% in harvested energy compared to the existing directly connected PV array scheme.

Games are energy-intensive applications on mobile devices. Optimizing the energy efficiency of games is hence critical for battery-limited mobile devices. Although the advent of energy-aware scheduling (EAS) integrated in recent devices has provided opportunities for improved energy management, the framework is not specifically tuned for game applications. In this paper, we aim to improve the energy efficiency of game applications running on EAS-enabled mobile devices. To this end, we first analyze the functional characteristics of games, and investigate the source of the energy inefficiency. We then propose a scheme, called System-level Energy-optimization for Game Applications (SEGA), to improve the energy efficiency of games. SEGA governs CPU and GPU power consumption in a tightly coupled manner by employing three key techniques: (1) Lsync-aware GPU DVFS governor, (2) adaptive capacity clamping, and (3) on-demand touch boosting. We implemented SEGA on the latest Android-based smartphones. The evaluation results for 23 popular games showed that SEGA reduced the energy consumption of the Google Pixel 2 XL and Samsung Galaxy S9 Plus smartphones, at the device level, by 6.1–22.3 and 4.0–11.7 percent, respectively, with a quality of service (QoS) degradation of 1.1 and 0.5 percent, on average.

Cell imbalance in a multicell battery occurs over time due to varying operating environments. This imbalance leads to overall inefficiency in battery discharging due to the relatively weak cells in the battery. Reconfiguring the cells in the battery is one option for addressing the problem, but relevant circuits may lead to severe safety issues. In this article, we aim to optimize the discharge efficiency of a multicell battery using safety-supplemented hardware. To this end, we first design a cell string-level reconfiguration scheme that is safe in hardware operations and also provides scalability due to the low switching complexity. Second, we propose a machine learning-based run-time switch control that considers various battery-related factors, such as the state of charge, state of health, temperature, and current distributions. Specifically, by exploiting the deep reinforcement learning (DRL) technique, we train the complex relationship among the battery factors and derive the best switch configuration in run-time. We implemented a hardware prototype, validated its functionalities, and evaluated the efficacy of the DRL-based control policy. The experimental results showed that the proposed scheme, along with the optimization method, improves the discharge efficiency of multicell batteries. In particular, the discharge efficiency gain is maximized when the cells constituting the battery are unevenly distributed in terms of cell health and exposed temperature.

The optimal use of batteries in drones is a critical issue for achieving both reliable operation and maximum flight time. The key is to acquire accurate information about the state of charge (SoC) of the battery in runtime. Drones typically employ series-connected lithium-ion polymer (Li-Po) battery cells, whose SoC is affected by many environmental factors as well as flight patterns. In this article, we propose a scheme, called Optrone, which maximizes the flight time of a drone while safely using the battery. Understanding the implications of the factors affecting the SoC of the drone's battery pack, we propose a three-level SoC, which is a metric for representing the SoC of a battery in runtime. We also provide various operating policies to users to improve the safety and efficiency of operating the drone. We implemented the prototype hardware and software for Optrone, and validated its operation in controlled and real environments. The experimental results in a controlled environment showed that the proposed three-level SoC poses less than 3% error and the operating policies achieved a flight time gain of 19.4%, while guaranteeing battery safety. We also observed a flight time gain of about 10% in real outdoor experiments, where the user rightly adheres to the advised Optrone policy.

Mobile users are always demanding extended availability in their battery use. Together with enlarged battery capacity, fast charging is one approach that provides an improved user experience in battery use. Recently, device manufacturers have been developing a variety of fast charging techniques for mobile devices. However, the existing techniques severely reduce the charging power when the device is in use while charging. We experimentally demonstrate that the primary cause of the reduction in the charge speed during device use is to cope with the performance degradation incurred by heat generation. We then propose an adaptive charging scheme, called Always-on Quick Charging, which enables fast charging especially when the device is in use. The key idea of our approach is to adjust the charging power while ensuring that the heat generated by the charging does not affect the performance. The proposed scheme is implemented in Google's Pixel 2XL smartphone. The experiment with a real-world usage scenario shows that the charging speed of the proposed scheme is up to 2.4 times faster than the default scheme, while preserving device performance.