Search

We prove that in asynchronous message-passing systems where at most one process may crash, there is no lock-free strongly linearizable implementation of a weak object that we call Test-or-Set (ToS). This object allows a single distinguished process to apply the set operation once, and a different distinguished process to apply the test operation also once. Since this weak object can be directly implemented by a single-writer single-reader (SWSR) register (and other common objects such as max-register, snapshot and counter), this result implies that there is no $1$-resilient lock-free strongly linearizable implementation of a SWSR register (and of these other objects) in message-passing systems. We also prove that there is no $1$-resilient lock-free \emph{write} strongly-linearizable implementation of a 2-writer 1-reader (2W1R) register in asynchronous message-passing systems., Comment: 12 pages

Policy gradient methods are appealing in deep reinforcement learning but suffer from high variance of gradient estimate. To reduce the variance, the state value function is applied commonly. However, the effect of the state value function becomes limited in stochastic dynamic environments, where the unexpected state dynamics and rewards will increase the variance. In this paper, we propose to replace the state value function with a novel hindsight value function, which leverages the information from the future to reduce the variance of the gradient estimate for stochastic dynamic environments. Particularly, to obtain an ideally unbiased gradient estimate, we propose an information-theoretic approach, which optimizes the embeddings of the future to be independent of previous actions. In our experiments, we apply the proposed hindsight value function in stochastic dynamic environments, including discrete-action environments and continuous-action environments. Compared with the standard state value function, the proposed hindsight value function consistently reduces the variance, stabilizes the training, and improves the eventual policy., Comment: Accepted by IJCAI2021

In a seminal work, Golab et al. showed that a randomized algorithm that works with atomic objects may lose some of its properties if we replace the atomic objects that it uses with linearizable objects. It was not known whether the properties that can be lost include the important property of termination (with probability 1). In this paper, we first show that, for randomized algorithms, termination can indeed be lost. Golab et al. also introduced strong linearizability, and proved that strongly linearizable objects can be used as if they were atomic objects, even for randomized algorithms: they preserve the algorithm's correctness properties, including termination. Unfortunately, there are important cases where strong linearizability is impossible to achieve. In particular, Helmi et al. MWMR registers do not have strongly linearizable implementations from SWMR registers. So we propose a new type of register linearizability, called write strong-linearizability, that is strictly stronger than linearizability but strictly weaker than strong linearizability. We prove that some randomized algorithms that fail to terminate with linearizable registers, work with write strongly-linearizable ones. In other words, there are cases where linearizability is not sufficient but write strong-linearizability is. In contrast to the impossibility result mentioned above, we prove that write strongly-linearizable MWMR registers are implementable from SWMR registers. Achieving write strong-linearizability, however, is harder than achieving just linearizability: we give a simple implementation of MWMR registers from SWMR registers and we prove that this implementation is linearizable but not write strongly-linearizable. Finally, we prove that any linearizable implementation of SWMR registers is necessarily write strongly-linearizable; this holds for shared-memory, message-passing, and hybrid systems.

We study the question of whether the "termination with probability 1" property of a randomized algorithm is preserved when one replaces the atomic registers that the algorithm uses with linearizable (implementations of) registers. We show that in general this is not so: roughly speaking, every randomized algorithm A has a corresponding algorithm A' that solves the same problem if the registers that it uses are atomic or strongly-linearizable, but does not terminate if these registers are replaced with "merely" linearizable ones. Together with a previous result shown in [15], this implies that one cannot use the well-known ABD implementation of registers in message-passing systems to automatically transform any randomized algorithm that works in shared-memory systems into a randomized algorithm that works in message-passing systems: with a strong adversary the resulting algorithm may not terminate.

Graph convolutional network (GCN) emerges as a promising direction to learn the inductive representation in graph data commonly used in widespread applications, such as E-commerce, social networks, and knowledge graphs. However, learning from graphs is non-trivial because of its mixed computation model involving both graph analytics and neural network computing. To this end, we decompose the GCN learning into two hierarchical paradigms: graph-level and node-level computing. Such a hierarchical paradigm facilitates the software and hardware accelerations for GCN learning. We propose a lightweight graph reordering methodology, incorporated with a GCN accelerator architecture that equips a customized cache design to fully utilize the graph-level data reuse. We also propose a mapping methodology aware of data reuse and task-level parallelism to handle various graphs inputs effectively. Results show that Rubik accelerator design improves energy efficiency by 26.3x to 1375.2x than GPU platforms across different datasets and GCN models.

Deep learning (DL) accelerators are increasingly deployed on edge devices to support fast local inferences. However, they suffer from a new security problem, i.e., being vulnerable to physical access based attacks. An adversary can easily obtain the entire neural network (NN) model by physically snooping the GDDR memory bus that connects the accelerator chip with DRAM memory. Therefore, memory encryption becomes important for DL accelerators on edge devices to improve the security of NN models. Nevertheless, we observe that traditional memory encryption solutions that have been efficiently used in CPU systems cause significant performance degradation when directly used in DL accelerators. The main reason comes from the big bandwidth gap between the GDDR memory bus and the encryption engine. To address this problem, our paper proposes SEAL, a Secure and Efficient Accelerator scheme for deep Learning. SEAL enhances the performance of the encrypted DL accelerator from two aspects, i.e., improving the data access bandwidth and the efficiency of memory encryption. Specifically, to improve the data access bandwidth, SEAL leverages a criticality-aware smart encryption scheme which identifies partial data that have no impact on the security of NN models and allows them to bypass the encryption engine, thus reducing the amount of data to be encrypted. To improve the efficiency of memory encryption, SEAL leverages a colocation mode encryption scheme to eliminate memory accesses from counters used for encryption by co-locating data and their counters. Our experimental results demonstrate that, compared with traditional memory encryption solutions, SEAL achieves 1.4 ~ 1.6 times IPC improvement and reduces the inference latency by 39% ~ 60%. Compared with a baseline accelerator without memory encryption, SEAL compromises only 5% ~ 7% IPC for significant security improvement., Comment: 14 pages, 15 figures

The well-known randomized consensus algorithm by Aspnes and Herlihy for asynchronous shared-memory systems was proved to work, even against a strong adversary, under the assumption that the registers that it uses are atomic registers. With atomic registers, every read or write operation is instantaneous (and thus indivisible). As pointed out by Golab et al. (2011), however, a randomized algorithm that works with atomic registers does not necessarily work if we replace the atomic registers that it uses with linearizable implementations of registers. This raises the following question: does the randomized consensus algorithm by Aspnes and Herlihy still work against a strong adversary if we replace its atomic registers with linearizable registers? We show that the answer is affirmative, in fact, we show that even linearizable registers are not necessary. More precisely, we prove that the algorithm by Aspnes and Herlihy works against a strong adversary even if the algorithm uses only regular registers.

In this work, we first characterize the hybrid execution patterns of GCNs on Intel Xeon CPU. Guided by the characterization, we design a GCN accelerator, HyGCN, using a hybrid architecture to efficiently perform GCNs. Specifically, first, we build a new programming model to exploit the fine-grained parallelism for our hardware design. Second, we propose a hardware design with two efficient processing engines to alleviate the irregularity of Aggregation phase and leverage the regularity of Combination phase. Besides, these engines can exploit various parallelism and reuse highly reusable data efficiently. Third, we optimize the overall system via inter-engine pipeline for inter-phase fusion and priority-based off-chip memory access coordination to improve off-chip bandwidth utilization. Compared to the state-of-the-art software framework running on Intel Xeon CPU and NVIDIA V100 GPU, our work achieves on average 1509$\times$ speedup with 2500$\times$ energy reduction and average 6.5$\times$ speedup with 10$\times$ energy reduction, respectively., Comment: To Appear in 2020 IEEE International Symposium on High Performance Computer Architecture

Recently, backpropagation through time inspired learning algorithms are widely introduced into SNNs to improve the performance, which brings the possibility to attack the models accurately given Spatio-temporal gradient maps. We propose two approaches to address the challenges of gradient input incompatibility and gradient vanishing. Specifically, we design a gradient to spike converter to convert continuous gradients to ternary ones compatible with spike inputs. Then, we design a gradient trigger to construct ternary gradients that can randomly flip the spike inputs with a controllable turnover rate, when meeting all zero gradients. Putting these methods together, we build an adversarial attack methodology for SNNs trained by supervised algorithms. Moreover, we analyze the influence of the training loss function and the firing threshold of the penultimate layer, which indicates a "trap" region under the cross-entropy loss that can be escaped by threshold tuning. Extensive experiments are conducted to validate the effectiveness of our solution. Besides the quantitative analysis of the influence factors, we evidence that SNNs are more robust against adversarial attack than ANNs. This work can help reveal what happens in SNN attack and might stimulate more research on the security of SNN models and neuromorphic devices.

The isospectral deformations of the Frobenius-Stickelberger-Thiele (FST) polynomials introduced in [32](Spiridonov et al. Commun. Math. Phys. 272:139--165, 2007 ) are studied. For a specific choice of the deformation of the spectral measure, one is led to an integrable lattice (FST lattice), which is indeed an isospectral flow connected with a generalized eigenvalue problem. In the second part of the paper the spectral problem used previously in the study of the modified Camassa-Holm (mCH) peakon lattice is interpreted in terms of the FST polynomials together with the associated FST polynomials, resulting in a map from the mCH peakon lattice to a negative flow of the finite FST lattice. Furthermore, it is pointed out that the degenerate case of the finite FST lattice unexpectedly maps to the interlacing peakon ODE system associated with the two-component mCH equation studied in [17](Chang et al. Adv. Math. 299:1--35, 2016)., Comment: 29 pages. online in Commun. Math. Phys

As well known, the huge memory and compute costs of both artificial neural networks (ANNs) and spiking neural networks (SNNs) greatly hinder their deployment on edge devices with high efficiency. Model compression has been proposed as a promising technique to improve the running efficiency via parameter and operation reduction. Whereas, this technique is mainly practiced in ANNs rather than SNNs. It is interesting to answer how much an SNN model can be compressed without compromising its functionality, where two challenges should be addressed: i) the accuracy of SNNs is usually sensitive to model compression, which requires an accurate compression methodology; ii) the computation of SNNs is event-driven rather than static, which produces an extra compression dimension on dynamic spikes. To this end, we realize a comprehensive SNN compression through three steps. First, we formulate the connection pruning and weight quantization as a constrained optimization problem. Second, we combine spatio-temporal backpropagation (STBP) and alternating direction method of multipliers (ADMM) to solve the problem with minimum accuracy loss. Third, we further propose activity regularization to reduce the spike events for fewer active operations. These methods can be applied in either a single way for moderate compression or a joint way for aggressive compression. We define several quantitative metrics to evaluation the compression performance for SNNs. Our methodology is validated in pattern recognition tasks over MNIST, N-MNIST, CIFAR10, and CIFAR100 datasets, where extensive comparisons, analyses, and insights are provided. To our best knowledge, this is the first work that studies SNN compression in a comprehensive manner by exploiting all compressible components and achieves better results., Comment: Under review