We introduce On-NAS, a memory-efficient on-device neural architecture search (NAS) solution, that enables memory-constrained embedded devices to find the best deep model architecture and train it on the device. Based ...
详细信息
ISBN:
(纸本)9798400704147
We introduce On-NAS, a memory-efficient on-device neural architecture search (NAS) solution, that enables memory-constrained embedded devices to find the best deep model architecture and train it on the device. Based on the cell-based differentiable NAS, it drastically curtails the massive memory requirement of architecture search, one of the major bottlenecks in realizing NAS on embedded devices. On-NAS first pre-trains a basic architecture block, called meta cell, by combining.. cells into a single condensed cell via two-fold meta-learning, which can flexibly evolve to various architectures, saving the device storage space.. times. Then, the offline-learned meta cell is loaded onto the device and unfolded to perform online on-device NAS via 1) expectation-based operation and edge pair search, enabling memory-efficient partial architecture search by reducing the required memory up to.. and../4 times, respectively, given.. candidate operations and.. nodes in a cell, and 2) step-by-step back-propagation that saves the memory usage of the backward pass of the..-cell architecture up to.. times. To the best of our knowledge, On-NAS is the firststandalone NAS and training solution fully operable on embedded devices with limited memory. Our experiment results show that On-NAS effectively identifies optimal architectures and trains it on the device, on par with GPU-based NAS in both few-shot and full-task learning settings, e.g., even 1.3% higher accuracy on miniImageNet, while reducing the run-time memory and storage usage up to 20x and 4x, respectively.
Continual Learning (CL) allows applications such as user personalization and household robots to learn on the fly and adapt to context. This is an important feature when context, actions, and users change. However, en...
详细信息
ISBN:
(纸本)9798400704147
Continual Learning (CL) allows applications such as user personalization and household robots to learn on the fly and adapt to context. This is an important feature when context, actions, and users change. However, enabling CL on resource-constrained embeddedsystems is challenging due to the limited labeled data, memory, and computing capacity. In this paper, we propose LifeLearner, a hardware-aware meta continual learning system that drastically optimizes system resources (lower memory, latency, energy consumption) while ensuring high accuracy. Specifically, we (1) exploit meta-learning and rehearsal strategies to explicitly cope with data scarcity issues and ensure high accuracy, (2) effectively combine lossless and lossy compression to significantly reduce the resource requirements of CL and rehearsal samples, and (3) developed hardware-aware system on embedded and IoT platforms considering the hardware characteristics. As a result, LifeLearner achieves near-optimal CL performance, falling short by only 2.8% on accuracy compared to an Oracle baseline. With respect to the state-of-the-art (SOTA) Meta CL method, LifeLearner drastically reduces the memory footprint ( by 178.7x), end-to-end latency by 80.8-94.2%, and energy consumption by 80.9-94.2%. In addition, we successfully deployed LifeLearner on two edge devices and a microcontroller unit, thereby enabling efficient CL on resource-constrained platforms where it would be impractical to run SOTA methods and the far-reaching deployment of adaptable CL in a ubiquitous manner. Code is available at https://***/theyoungkwon/LifeLearner.
We are demonstrating a networking technology and application environment that connects highly-constrained low-power wireless embeddedsensor networks with large-scale IP networks. This technology is based on the 6LoWP...
详细信息
ISBN:
(纸本)9781595937636
We are demonstrating a networking technology and application environment that connects highly-constrained low-power wireless embeddedsensor networks with large-scale IP networks. This technology is based on the 6LoWPAN IPv6-over-802.15.4 adptation layer.
This paper considers the quantized H-infinity filtering problem of networkedsystems with random sensor delays. A quantized random delay model is adopted to study the relationship of the communication delay, quantizat...
详细信息
ISBN:
(纸本)9781424427239
This paper considers the quantized H-infinity filtering problem of networkedsystems with random sensor delays. A quantized random delay model is adopted to study the relationship of the communication delay, quantization and the system performance. The resulting design guarantees that the error system exponentially mean-square stable and with a prescribed H-infinity performance bound. A numerical example is given to illustrated the effectiveness of the proposed H-infinity filtering strategy.
Audio is valuable in many mobile, embedded, and cyber-physical systems. We propose AvA, an acoustic adaptive filtering architecture, configurable to a wide range of applications and systems. By incorporating AvA into ...
详细信息
ISBN:
(纸本)9781665496247
Audio is valuable in many mobile, embedded, and cyber-physical systems. We propose AvA, an acoustic adaptive filtering architecture, configurable to a wide range of applications and systems. By incorporating AvA into their own systems, developers can select which sounds to enhance or filter out depending on their application needs. AvA accomplishes this by using a novel adaptive beamforming algorithm called content-informed adaptive beamforming (CIBF), that directly uses detectors and sound models that developers have created for their own applications to enhance or filter out sounds. CIBF uses a novel three step approach to propagate gradients from a wide range of different model types and signal feature representations to learn filter coefficients. We apply AvA to four scenarios and demonstrate that AvA enhances their respective performances by up to 11.1%. We also integrate AvA into two different mobile/embedded platforms with widely different resource constraints and target sounds/noises to show the boosts in performance and robustness these applications can see using AvA.
The proceedings contain 35 papers. The topics discussed include: FWB: funneling wider bandwidth algorithm for high performance data collection in wireless sensor networks;balancing energy harvesting and transmission s...
ISBN:
(纸本)9781450359603
The proceedings contain 35 papers. The topics discussed include: FWB: funneling wider bandwidth algorithm for high performance data collection in wireless sensor networks;balancing energy harvesting and transmission scheduling in aggregation convergecast;routing with renewable energy management in wireless sensor networks;network alarm flood pattern mining algorithm based on multi-dimensional association;constructing an accurate and a high-performance power profiler for embeddedsystems and smartphones;maximizing mobiles energy saving through tasks optimal offloading placement in two-tier cloud;and delay-sensitive multiplayer augmented reality game planning in mobile edge computing.
Wireless sensor Networks (WSN) promise researchers a powerful instrument for observing sizable phenomena with fine granularity over long periods. Since the accuracy of data is important to the whole system's perfo...
详细信息
ISBN:
(纸本)9781605587486
Wireless sensor Networks (WSN) promise researchers a powerful instrument for observing sizable phenomena with fine granularity over long periods. Since the accuracy of data is important to the whole system's performance, detecting nodes with faulty readings is an essential issue in network management. As a complementary solution to detecting nodes with functionnal faults, this paper proposes FIND, a novel method to detect nodes with data faults that neither assumes a particular sensing model nor requires costly event injections. After the nodes in a network detect a natural event, FIND ranks the nodes based on their sensing readings as well as their physical distances from the event. FIND works for systems where the measured signal attenuates with distance. A node is considered faulty if there is a significant mismatch between the sensor data rank and the distance rank Theoretically, we show that average ranking difference is a provable indicator of possible data faults. FIND is extensively evaluated in simulations and two test bed experiments with up to 25 MicaZ nodes. Evaluation shows that FIND has a less than 5% miss detection rate and false alarm rate in most noisy environments.
This paper demonstrates the implementation of TinyTune, a collaborative musical instrument using sensor motes. The system implementation is distributed across multiple nodes and supports the basic elements of a musica...
详细信息
ISBN:
(纸本)9781450303446
This paper demonstrates the implementation of TinyTune, a collaborative musical instrument using sensor motes. The system implementation is distributed across multiple nodes and supports the basic elements of a musical instrument, such as pitch and octave selection. The communication design for realizing a collaborative musical instrument and the available user configuration options are then presented. Other topics of discussion include: the underlying system architecture, covering the advantages of our design choices;and the extensibility of the concept, which discusses how the nodes are configured in multi-instrument environments.
暂无评论