Many computer vision based mobile applications, ranging from simple children's games to complex, safety-critical applications use convolutional neural networks (CNNs) to provide accurate image recognition of both still images and live video streams. Current state-of-the-art CNNs, however, are very computationally complex and not well suited to real-time mobile computation, where computing resources are severely limited. Current work to improve the latency of CNNs on mobile devices include model compression, caching, as well as CNNs designed specifically for mobile use (e.g. MobileNet). All of these approaches, however, suffer from significantly reduced accuracy. In this paper we introduce a class-aware caching scheme for CNNs that provides fast, high-quality image recognition while not sacrificing accuracy. During offline computation, our system learns the most valuable filters for each class of a dataset through our proposed innovative filter ranking algorithm. At inference time we quickly determine the top-5 most likely classifications of an input image using an LSH, which is very fast but not as accurate as a CNN. Then, using the filter rankings we compute the valuable filters for the top-5 predicted classes while re-using previous computations for any others. This allows the CNN to maintain flexibility, and robustness which are often lost with model compression or mobile-based CNNs, while still providing significant latency reduction. Our evaluation shows that we can achieve latency reduction of up to 30 % while providing a similar level of accuracy.

Computer vision on smartphones is commonly achieved through the use of convolutional neural networks (CNNs). CNNs offer accurate image recognition, but struggle with latency when run with resource constraints. Current work in mobile CNNs aim to improve the latency on mobile devices through techniques such as quantization, model compression, early-exit, caching, etc. While these methods can improve the overall latency of image recognition, they also sacrifice significant accuracy. This problem is compounded by the wide range of pre-trained CNNs that have been released over the past decade. Many of these useful CNNs were developed before innovations that improved the efficiency of CNNs. Anyone attempting to use a CNN trained many years ago may be out of luck.In this paper, we introduce a computation caching scheme paired with early-exit strategies to improve the latency of CNNs on smartphones. Our system has both offline and online components. The offline system is used to find patterns in CNN execution, which are stored on the device. The online component reviews the current state of a CNN execution, determines if it matches a saved pattern, and can choose to forgo the full CNN execution and return a classification immediately. This system improves the speed of image recognition on any smartphone. It also requires no modification to the CNN, and no CNN training, meaning that it can be applied to any CNN that you can find, including many developed years ago. Our system achieves an average latency reduction of 17% while maintaining strong accuracy.

Optical wireless communication (OWC) shows great potential due to its broad spectrum and the exceptional intensity switching speed of LEDs. Under poor conditions, most OWC systems switch from complex and more error prone high-order modulation schemes to more robust On-Off Keying (OOK) modulation defined in the IEEE OWC standard. This paper presents LiFOD, a high-speed indoor OOK-based OWC system with fine-grained dimming support. While ensuring fine-grained dimming, LiFOD remarkably achieves robust communication at up to 400 Kbps at a distance of 6 meters. This is the first time that the data rate has improved via OWC dimming in comparison to the previous approaches that consider trading off dimming and communication. LiFOD makes two key technical contributions. First, LiFOD utilizes Compensation Symbols (CS) as a reliable side-channel to represent bit patterns dynamically and improve throughput. We firstly design greedy-based bit pattern mining. Then we propose 2D feature enhancement via YOLO model for real-time bit pattern mining. Second, LiFOD synchronously redesigns optical symbols and CS relocation schemes for fine-grained dimming and robust decoding. Experiments on low-cost Beaglebone prototypes with commercial LED lamps and the photodiode (PD) demonstrate that LiFOD significantly outperforms the state-of-the-art system with 2.1× throughput on the SIGCOMM17 data-trace.

Optical Camera Communication (OCC) appears as a promising technology to provide secure and pervasive wireless services with users' daily smart devices. Rolling shutter based modulations can improve the frequency response of the camera. This paper introduces a 2D Rolling Block (2DRB) based OCC modulation to use un-exploited spatial diversity to improve OCC's data rate for real-world applications. 2DRB outperforms traditional 1D strip based modulations. Using our 2DRB prototype with commercial devices, we show a significant data rate enhancement. We also discuss one promising real-world use case: indoor office integrated lighting and communication.

Underwater optical wireless communication techniques are promising due to a broad bandwidth with a long communication range compared with existing expensive acoustic and RF-based underwater communication techniques. For underwater navigation assistance during dive and rescue, it is more practical to adopt passive optical tags for objects/human identification and location-based services. However, existing optical tags (bar/QR codes) employ one/two dimensional designs, which lack significant element/symbol distance for robust decoding and full-directional localization capabilities for underwater navigation tasks. This paper investigates opportunities to increase the element distance in passive low-order optical tags by exploiting 3D spatial diversity. Specifically, we design U-Star, a system that consists of Underwater Optical Identification (UOID) tags and commercial camera-based tag readers for underwater navigation. Our UOID tags embed rich location and guidance information. Additionally, because our UOID tags employ a three-dimensional design, they can also determine the relative location of a user in real-time based on the perspective principles. We design AI based mobile algorithms for underwater denoising, relative positioning, and robust data parsing for tag readers. Finally, we evaluate U-Star on real UOID tag prototypes under different underwater scenarios. Results show that our 3-order UOID tag can embed 21 bits with a BER of 0.003 at 1m and less than 0.05 at up to 3m, which is sufficient for underwater navigation guidance with backup database.

Many emerging multimedia mobile applications rely heavily upon image recognition of both static images and live video streams. Image recognition is commonly achieved using deep neural networks (DNNs) which can achieve high accuracy but also incur significant computation latency and energy con-sumption on resource-constrained smartphones. Recent efforts addressing these issues include cloud offloading and reducing the complexity of the DNNs, which, however, introduce increased network latency or reduced accuracy. In-memory caching has also been explored to assess the similarity of images as opposed to exact matching. However, such approximate caching systems often treat devices as static nodes, and do not fully utilize the mobile and collaborative nature of smartphones without outside infrastructure. Another consequence of treating nodes as static is the necessity of cache sizes larger than what is feasible for individual mobile applications. In this paper we introduce Co-Cache, a in-memory caching paradigm that supports infrastructure-less collaborative compu-tation reuse in smartphone image recognition. Co-Cache utilizes the inertial movement of smartphones, the locality inherent in video streams, as well as information from nearby, peer-to-peer devices to maximize the computation reuse opportunities in mobile image recognition. Compared to other caching systems, our extensive evaluation shows that Co-Cache can reduce the required number of cache entries by 50–70 % while lowering the average latency of standard image recognition applications by up to 94 % with minimal loss of recognition accuracy.

Optical wireless communication (OWC) shows great potential for high-speed communication due to its broad spectrum and the exceptional intensity switching speed of LEDs. Under poor conditions, most OWC systems switch from complex and more error prone high-order modulation schemes to the more robust On-Off Keying (OOK) modulation defined in the IEEE OWC standard. This paper presents LiFOD, a high-speed indoor OOK-based OWC system with fine-grained dimming support. While ensuring fine-grained dimming, LiFOD remark-ably achieves robust communication at up to 400 Kbps at a distance of 6 meters. This is the first time that the data rate has improved via OWC dimming in comparison to the previous approaches that consider trading off dimming and communication. LiFOD makes two key technical contributions. First, LiFOD utilizes Compensation Symbols (CS) as a reliable side-channel to represent bit patterns dynamically and improve throughput. Second, LiFOD synchronously redesigns optical symbols and CS relocation schemes for fine-grained dimming and robust decoding. Experiments on low-cost Beaglebone prototypes with commercial LED lamps and the photodiode (PD) demonstrate that LiFOD significantly outperforms the state-of-art system with at least 2.1x throughput on the SIGCOMM17 data-trace.

Many emerging mobile applications rely heavily upon image recognition of both static images and live video streams. Image recognition is commonly achieved using deep neural networks (DNNs) which can achieve high accuracy but also incur significant computation latency and energy consumption on resource-constrained smartphones. We introduce an in-memory caching paradigm that supports infrastructure-less collaborative computation reuse in smartphone image recognition. We propose using the inertial movement of smartphones, the locality inherent in video streams, as well as information from nearby, peer-to-peer devices to maximize the computation reuse opportunities in mobile image recognition. Experimental results show that our system lowers the average latency of standard mobile neural network image recognition applications by up to 94% with minimal loss of recognition accuracy.

In recent years mobile computing has been rapidly expanding to the point that there are now more devices than there are people. While once it was common for every household to have one PC, it is now common for every person to have a mobile device. With the increased use of smartphone devices, there has also been an increase in the need for mobile ad hoc networks, in which phones connect directly to each other without the need for an intermediate router. Most modern smart phones are equipped with both Bluetooth and Wifi Direct, where Wifi Direct has a better transmission range and rate and Bluetooth is more energy efficient. However only one or the other is used in a smartphone ad hoc network. We propose HSNet, a framework to enable the automatic switching between Wifi Direct and Bluetooth to emphasize minimizing energy consumption while still maintaining an efficient network. We develop an application to evaluate the HSNet framework which shows significant energy savings when utilizing our switching algorithm to send messages by a less energy intensive technology in situations where energy conservation is desired. We discuss additional features of HSNet such as load balancing to help increase the lifetime of the network by more evenly distributing slave nodes among connected master nodes. Finally, we show that the throughput of our system is not affected due to technology switching for most scenarios.