This paper presents AppealNet, a novel edge/cloud collaborative architecture that runs deep learning (DL) tasks more efficiently than state-of-the-art solutions. For a given input, AppealNet accurately predicts on-the-fly whether it can be successfully processed by the DL model deployed on the resource-constrained edge device, and if not, appeals to the more powerful DL model deployed at the cloud. This is achieved by employing a two-head neural network architecture that explicitly takes inference difficulty into consideration and optimizes the tradeoff between accuracy and computation/communication cost of the edge/cloud collaborative architecture. Experimental results on several image classification datasets show up to more than 40% energy savings compared to existing techniques without sacrificing accuracy.

Deep neural networks (DNNs) have become one of the enabling technologies in many safety-critical applications, e.g., autonomous driving and medical image analysis. DNN systems, however, suffer from various kinds of threats, such as adversarial example attacks and fault injection attacks. While there are many defense methods proposed against maliciously crafted inputs, solutions against faults presented in the DNN system itself (e.g., parameters and calculations) are far less explored. In this paper, we develop a novel lightweight fault-tolerant solution for DNN-based systems, namely DeepDyve, which employs pre-trained neural networks that are far simpler and smaller than the original DNN for dynamic verification. The key to enabling such lightweight checking is that the smaller neural network only needs to produce approximate results for the initial task without sacrificing fault coverage much. We develop efficient and effective architecture and task exploration techniques to achieve optimized risk/overhead trade-off in DeepDyve. Experimental results show that DeepDyve can reduce 90% of the risks at around 10% overhead.

Deep Neural Networks (DNNs) have attracted mainstream adoption in various application domains. Their reliability and security are therefore serious concerns in those safety-critical applications such as surveillance and medical systems. In this paper, we propose a novel dual modular redundancy framework for DNNs, namely D2NN, which is able to tradeoff the system robustness with overhead in a fine-grained manner. We evaluate D2NN framework with DNN models trained on MNIST and CIFAR10 datasets under fault injection attacks, and experimental results demonstrate the efficacy of our proposed solution.

Computation kernels in emerging recognition, mining, and synthesis (RMS) applications are inherently error-resilient, where approximate computing can be applied to improve their energy efficiency by trading off computational effort and output quality. One promising approximate computing technique is to perform approximate computing with memory, which stores a subset of function responses in a lookup table (LUT), and avoids redundant computation when encountering similar input patterns. Limited by the memory space, most existing solutions simply store values for those frequently-appeared input patterns, without considering output quality and/or intrinsic characteristic of the target kernel. In this paper, we propose a novel LUT allocation technique for approximate computing with memory, which is able to dramatically improve the hit rate of LUT and hence achieves significant energy savings under given quality constraints. We also present how to apply the proposed LUT allocation solution for multiple computation kernels. Experimental results show the efficacy of our proposed methodology.

Energy harvesting systems recycling energy from ambient environment for extended lifetime have been proposed to be the next step in the evolution of the Internet of Things. One of the largest challenges of energy harvesting is the insufficiency and instability of ambient energy sources. Unlike wireless distributed systems considered to date, frequent power failures need to be taken into account in designing energy-efficient transmission strategies. In this paper, we propose a message passing method between energy harvesting nodes, which predicts the probability of success before message passing and adjusts transmission strategy online according to input power trace. Experimental result shows that our method can reduce energy consumption of transmission significantly compared to traditional ones and keeps effective for various kinds of unstable input power.

Computing with memory, which stores function responses of some input patterns into lookup tables offline and retrieves their values when encountering similar patterns (instead of performing online calculation), is a promising energy-efficient computing technique. No doubt to say, with a given lookup table size, the efficiency of this technique depends on which function responses are stored and how they are organized. In this paper, we propose a novel adaptive approximate lookup table based accelerator, wherein we store function responses in a hierarchical manner with increasing fine-grained granularity and accuracy. In addition, the proposed accelerator provides lightweight compensation on output results at different precision levels according to input patterns and output quality requirements. Moreover, our accelerator conducts adaptive lookup table search by exploiting input locality. Experimental results on various computation kernels show significant energy savings of the proposed accelerator over prior solutions.

Many emerging applications are inherently error-resilient and hence do not require exact computation. In this paper, we consider the task allocation and scheduling problem for mapping such applications to voltage-scalable multiprocessor systems. The proposed solution, namely ApproxMap, judiciously determines the mapping and execution sequence of resilient tasks to minimize the energy consumption of the application while meeting their target quality requirements and timing constraints. To be specific, ApproxMap generates energy-efficient yet flexible task schedule at design-time, and conducts lightweight online adjustment according to runtime dynamics for further energy-efficiency improvement. Experimental results on various task graphs demonstrate the efficacy of ApproxMap.

Motivated by the inherent error-resilience of emerging recognition, mining, and synthesis (RMS) applications, approximate computing techniques such as precision scaling has been advocated for achieving energy-efficiency gains at the cost of small accuracy loss. Most existing solutions, however, focus on the approximation of on-chip computations without considering that of off-chip data accesses, whose energy consumption may contribute to a significant portion of the total energy. In this work, we propose a novel approximate memory access technique for dynamic precision scaling, namely ApproxMA. To be specific, by taking both runtime data precision constraints and error-resilient capabilities of the application into consideration, ApproxMA determines the precision of data accesses and loads scaled data from off-chip memory for computation. Experimental results with mixture model-based clustering algorithms demonstrate the efficacy of the proposed methodology.

Recognition, Mining, and Synthesis (RMS) applications are expected to make up much of the computing workloads of the future. Many of these applications (e.g., recommender systems and search engine) are formulated as finding eigenvalues/vectors of large-scale matrices. These applications are inherently error-tolerant, and it is often unnecessary, sometimes even impossible, to calculate all the eigenpairs. Motivated by the above, in this work, we propose a novel approximate computing technique for large-scale eigen-decomposition, namely ApproxEigen, wherein we focus on the practically-used Krylov subspace methods to find finite number of eigenpairs. With ApproxEigen, we provide a set of computation kernels with different levels of approximation for data pre-processing and solution finding, and conduct accuracy tuning under given quality constraints. Experimental results demonstrate that ApproxEigen is able to achieve significant energy-efficiency improvement while keeping high accuracy.