Search

Time series foundation models have demonstrated strong performance in zero-shot learning, making them well-suited for predicting rapidly evolving patterns in real-world applications where relevant training data are scarce. However, most of these models rely on the Transformer architecture, which incurs quadratic complexity as input length increases. To address this, we introduce TSMamba, a linear-complexity foundation model for time series forecasting built on the Mamba architecture. The model captures temporal dependencies through both forward and backward Mamba encoders, achieving high prediction accuracy. To reduce reliance on large datasets and lower training costs, TSMamba employs a two-stage transfer learning process that leverages pretrained Mamba LLMs, allowing effective time series modeling with a moderate training set. In the first stage, the forward and backward backbones are optimized via patch-wise autoregressive prediction; in the second stage, the model trains a prediction head and refines other components for long-term forecasting. While the backbone assumes channel independence to manage varying channel numbers across datasets, a channel-wise compressed attention module is introduced to capture cross-channel dependencies during fine-tuning on specific multivariate datasets. Experiments show that TSMamba's zero-shot performance is comparable to state-of-the-art time series foundation models, despite using significantly less training data. It also achieves competitive or superior full-shot performance compared to task-specific prediction models. The code will be made publicly available.

In this study,we investigate the characteristics of three-dimensional turbulent boundary layers influenced by transverse flow and pressure gradients. Our findings reveal that even without assuming an infinite sweep, a fully developed turbulent boundary layer over the present swept blunt body maintains spanwise homogeneity, consistent with infinite sweep assumptions.We critically examine the law-of-the and temperature-velocity relationships, typically applied two-dimensional turbulent boundary layers, in three-dimensional contexts. Results show that with transverse velocity and pressure gradient, streamwise velocity adheres to classical velocity transformation relationships and the predictive accuracy of classical temperaturevelocity relationships diminishes because of pressure gradient. We show that near-wall streak structures persist and correspond with energetic structures in the outer region, though three-dimensional effects redistribute energy to align more with the external flow direction. Analysis of shear Reynolds stress and mean flow shear directions reveals in near-wall regions with low transverse flow velocity, but significant deviations at higher transverse velocities. Introduction of transverse pressure gradients together with the transverse velocities alter the velocity profile and mean flow shear directions, with shear Reynolds stress experiencing similar changes but with a lag increasing with transverse. Consistent directional alignment in outer regions suggests a partitioned relationship between shear Reynolds stress and mean flow shear: nonlinear in the inner region and approximately linear in the outer region.

The attachment-line boundary layer is critical in hypersonic flows because of its significant impact on heat transfer and aerodynamic performance. In this study, high-fidelity numerical simulations are conducted to analyze the subcritical roughness-induced laminar-turbulent transition at the leading-edge attachment-line boundary layer of a blunt swept body under hypersonic conditions. This simulation represents a significant advancement by successfully reproducing the complete leading-edge contamination process induced by surface roughness elements in a realistic configuration, thereby providing previously unattainable insights. Two roughness elements of different heights are examined. For the lower-height roughness element, additional unsteady perturbations are required to trigger a transition in the wake, suggesting that the flow field around the roughness element acts as a disturbance amplifier for upstream perturbations. Conversely, a higher roughness element can independently induce the transition. A low-frequency absolute instability is detected behind the roughness, leading to the formation of streaks. The secondary instabilities of these streaks are identified as the direct cause of the final transition.

With current and future leading systems adopting heterogeneous architectures, adapting existing models for heterogeneous supercomputers is of urgent need for improving model resolution and reducing modeling uncertainty. This paper presents our three-week effort on porting a complex earth system model, CESM 2.2, to a 40-million-core Sunway supercomputer. Taking a non-intrusive approach that tries to minimizes manual code modifications, our project tries to achieve both improvement of performance and consistency of the model code. By using a hierarchical grid system and an OpenMP-based offloading toolkit, our porting and parallelization effort covers over 80% of the code, and achieves a simulation speed of 340 SDPD (simulated days per day) for 5-km atmosphere, 265 SDPD for 3-km ocean, and 222 SDPD for a coupled model, thus making multi-year or even multi-decadal experiments at such high resolution possible., Comment: 18 pages, 13 figures

We investigate the relaxation dynamics of the fermion two-point correlation function $C_{mn}(t)=\langle\psi(t)|c_{m}^{\dag}c_{n}|\psi(t)\rangle$ in the XY chain with staggered nearest-neighbor hopping interaction after a quench. We find that the deviation $\delta C_{mn}(t)=C_{mn}(t)-C_{mn}(\infty)$ decays with time following the power law behavior $t^{-\mu}$, where the exponent $\mu$ depends on whether the quench is to the commensurate phase ($\mu=1$) and incommensurate phase ($\mu=\frac{1}{2}$). This decay of $\delta C_{mn}(t)$ arises from the transient behavior of the double excited quasiparticle occupations and the transitions between different excitation spectra. Furthermore, we find that the steady value $C_{mn}(\infty)$, which is different from the ground state expectation value, only involves the average fermion occupation numbers (i.e. the average excited single particle). We also observe nonanalytic singularities in the steady value $C_{mn}(\infty)$ for the quench to the critical points of the quantum phase transitions (QPTs), suggesting its potential use as a signature of QPTs., Comment: 12 pages, 8 figures

Large, high-capacity models trained on diverse datasets have shown remarkable successes on efficiently tackling downstream applications. In domains from NLP to Computer Vision, this has led to a consolidation of pretrained models, with general pretrained backbones serving as a starting point for many applications. Can such a consolidation happen in robotics? Conventionally, robotic learning methods train a separate model for every application, every robot, and even every environment. Can we instead train generalist X-robot policy that can be adapted efficiently to new robots, tasks, and environments? In this paper, we provide datasets in standardized data formats and models to make it possible to explore this possibility in the context of robotic manipulation, alongside experimental results that provide an example of effective X-robot policies. We assemble a dataset from 22 different robots collected through a collaboration between 21 institutions, demonstrating 527 skills (160266 tasks). We show that a high-capacity model trained on this data, which we call RT-X, exhibits positive transfer and improves the capabilities of multiple robots by leveraging experience from other platforms. More details can be found on the project website https://robotics-transformer-x.github.io., Comment: Project website: https://robotics-transformer-x.github.io

The next generation Sunway supercomputer employs the SW26010pro processor, which features a specialized on-chip heterogeneous architecture. Applications with significant hotspots can benefit from the great computation capacity improvement of Sunway many-core architectures by carefully making intensive manual many-core parallelization efforts. However, some legacy projects with large codebases, such as CESM, ROMS and WRF, contain numerous lines of code and do not have significant hotspots. The cost of manually porting such applications to the Sunway architecture is almost unaffordable. To overcome such a challenge, we have developed a toolkit named O2ATH. O2ATH forwards GNU OpenMP runtime library calls to Sunway's Athread library, which greatly simplifies the parallelization work on the Sunway architecture.O2ATH enables users to write both MPE and CPE code in a single file, and parallelization can be achieved by utilizing OpenMP directives and attributes. In practice, O2ATH has helped us to port two large projects, CESM and ROMS, to the CPEs of the next generation Sunway supercomputers via the OpenMP offload method. In the experiments, kernel speedups range from 3 to 15 times, resulting in 3 to 6 times whole application speedups.Furthermore, O2ATH requires significantly fewer code modifications compared to manually crafting CPE functions.This indicates that O2ATH can greatly enhance development efficiency when porting or optimizing large software projects on Sunway supercomputers., Comment: 15 pages, 6 figures, 5 tables

We investigate the dynamical relaxation behavior of the two-point correlation in extended XY models with a gapless phase after quenches from various initial states. Specifically, we study the XY chain with gapless phase induced by the additional interactions: Dzyaloshinskii-Moriya interaction and XZY-YZX type of three-site interaction. When quenching from the gapped phase, we observe that the additional interactions have no effect on the relaxation behavior. The relaxation behavior is $\delta C_{mn}(t)\sim t^{-3/2}$ and $\sim t^{-1/2}$ for the quench to the commensurate phase and the incommensurate phase, respectively. However, when quenching from the gapless phase, we demonstrate that the scaling behavior of $\delta C_{mn}(t)$ is changed to $\sim t^{-1}$ for the quench to the commensurate phase, and the decay of $\delta C_{mn}(t)$ follows $\sim t^{-1}$ or $\sim t^{-1/2}$ for the quench to the incommensurate phase depending on the parameters of pre-quench Hamiltonian. We also establish the dynamical phase diagrams based on the dynamical relaxation behavior of $\delta C_{mn}(t)$ in the extended XY models., Comment: 12 pages, 10 figures

A state-of-the-art large eddy simulation code has been developed to solve compressible flows in turbomachinery. The code has been engineered with a high degree of scalability, enabling it to effectively leverage the many-core architecture of the new Sunway system. A consistent performance of 115.8 DP-PFLOPs has been achieved on a high-pressure turbine cascade consisting of over 1.69 billion mesh elements and 865 billion Degree of Freedoms (DOFs). By leveraging a high-order unstructured solver and its portability to large heterogeneous parallel systems, we have progressed towards solving the grand challenge problem outlined by NASA, which involves a time-dependent simulation of a complete engine, incorporating all the aerodynamic and heat transfer components., Comment: SC23, November, 2023, Denver, CO., USA

Existing large language models have to run K times to generate a sequence of K tokens. In this paper, we present RecycleGPT, a generative language model with fast decoding speed by recycling pre-generated model states without running the whole model in multiple steps. Our approach relies on the observation that adjacent tokens in a sequence usually have strong correlations and the next token in a sequence can be reasonably guessed or inferred based on the preceding ones. Experiments and analysis demonstrate the effectiveness of our approach in lowering inference latency, achieving up to 1.4x speedup while preserving high performance., Comment: Technical Report

In one-dimensional (1D) disorder-free interacting systems, a sufficiently strong linear potential can induce localization of the many-body eigenstates, a phenomenon dubbed as Stark many-body localization (MBL). In this paper, we investigate the fate of Stark MBL in 1D spinless fermions systems with long-range interactions, specifically focusing on the role of interaction strength. We obtain the Stark MBL phase diagrams by computing the mean gap ratio and many-body inverse participation ratio at half-filling. We show that, for short-range interactions, there is a qualitative symmetry between the limits of weak and strong interactions. However, this symmetry is absent in the case of long-range interactions, where the system is always Stark many-body localized at strong interactions, regardless of the linear potential strength. Furthermore, we study the dynamics of imbalance and entanglement with various initial states using time-dependent variational principle (TDVP) numerical methods. We reveal that the dynamical quantities display a strong dependence on the initial conditions, which suggests that the Hilbert-space fragmentation precludes thermalization. Our results demonstrate the robustness of Stark MBL even in the presence of long-range interactions and offer an avenue to explore MBL in disorder-free systems with long-range interactions.

Anderson mixing (AM) is a classical method that can accelerate fixed-point iterations by exploring historical information. Despite the successful application of AM in scientific computing, the theoretical properties of AM are still under exploration. In this paper, we study the restarted version of the Type-I and Type-II AM methods, i.e., restarted AM. With a multi-step analysis, we give a unified convergence analysis for the two types of restarted AM and justify that the restarted Type-II AM can locally improve the convergence rate of the fixed-point iteration. Furthermore, we propose an adaptive mixing strategy by estimating the spectrum of the Jacobian matrix. If the Jacobian matrix is symmetric, we develop the short-term recurrence forms of restarted AM to reduce the memory cost. Finally, experimental results on various problems validate our theoretical findings.

In order to enhance the performance of Transformer models for long-term multivariate forecasting while minimizing computational demands, this paper introduces the Joint Time-Frequency Domain Transformer (JTFT). JTFT combines time and frequency domain representations to make predictions. The frequency domain representation efficiently extracts multi-scale dependencies while maintaining sparsity by utilizing a small number of learnable frequencies. Simultaneously, the time domain (TD) representation is derived from a fixed number of the most recent data points, strengthening the modeling of local relationships and mitigating the effects of non-stationarity. Importantly, the length of the representation remains independent of the input sequence length, enabling JTFT to achieve linear computational complexity. Furthermore, a low-rank attention layer is proposed to efficiently capture cross-dimensional dependencies, thus preventing performance degradation resulting from the entanglement of temporal and channel-wise modeling. Experimental results on six real-world datasets demonstrate that JTFT outperforms state-of-the-art baselines in predictive performance.

We report new Gaussian boson sampling experiments with pseudo-photon-number-resolving detection, which register up to 255 photon-click events. We consider partial photon distinguishability and develop a more complete model for the characterization of the noisy Gaussian boson sampling. In the quantum computational advantage regime, we use Bayesian tests and correlation function analysis to validate the samples against all current classical mockups. Estimating with the best classical algorithms to date, generating a single ideal sample from the same distribution on the supercomputer Frontier would take ~ 600 years using exact methods, whereas our quantum computer, Jiuzhang 3.0, takes only 1.27 us to produce a sample. Generating the hardest sample from the experiment using an exact algorithm would take Frontier ~ 3.1*10^10 years., Comment: PRL 2023 to appear

We investigate the dynamical quantum phase transition (DQPT) in the multi-band Bloch Hamiltonian of the one-dimensional periodic Kitaev model, focusing on quenches from a Bloch band. By analyzing the dynamical free energy and Pancharatnam geometric phase, we show that the critical times of DQPTs deviate from periodic spacing due to the multi-band effect, contrasting with results from two-band models. We propose a geometric interpretation to explain this non-uniform spacing. Additionally, we clarify the conditions needed for DQPT occurrence in the multi-band Bloch Hamiltonian, highlighting that a DQPT only arises when the quench from the Bloch states collapses the band gap at the critical point. Moreover, we establish that the dynamical topological order parameter, defined by the winding number of the Pancharatnam geometric phase, is not quantized but still exhibits discontinuous jumps at DQPT critical times due to periodic modulation. Additionally, we extend our analysis to mixed-state DQPT and find its absence at non-zero temperatures., Comment: 11 pages, 11 figures

The quantum supremacy experiment, such as Google Sycamore [Nature \textbf{574}, 505 (2019)], poses great challenge for classical verification due to the exponentially-increasing compute cost. Using a new-generation Sunway supercomputer within $8.5$ days, we provide a direct verification by computing three million exact amplitudes for the experimentally generated bitstrings, obtaining an XEB fidelity of $0.191\%$ (the estimated value is $0.224\%$). The leap of simulation capability is built on a multiple-amplitude tensor network contraction algorithm which systematically exploits the ``classical advantage" (the inherent ``store-and-compute" operation mode of von Neumann machines) of current supercomputers, and a fused tensor network contraction algorithm which drastically increases the compute efficiency on heterogeneous architectures. Our method has a far-reaching impact in solving quantum many-body problems, statistical problems as well as combinatorial optimization problems., Comment: 5 pages, 3 figures, comments are welcome!

The dynamics of the geometric phase are studied in inhomogeneous quantum spin chains after a quench. Analytic expressions of the Pancharatnam geometric phase (PGP) $\mathcal{G}(t)$ are derived, for both the period-two quantum Ising chain (QIC) and the disordered QIC. In the period-two QIC, due to the periodic modulation, the PGP changes with time at the boundary of the Brillouin zone, and consequently, the winding number $\nu_{D}(t)=\int_{0}^{\pi}[\partial\phi_{k}^{G}(t)/\partial k]dk/2\pi$ based on the PGP is not quantized and thus not topological anymore. Nevertheless, the PGP and its winding number show non-analytic singularities at the critical times of the dynamical quantum phase transitions (DQPTs). This relation between the PGP and the DQPT is further confirmed in the disordered QIC, where the winding number is not defined. It is found that the critical time of DQPT inherited from the homogeneous system and the additional one induced by the weak disorder are also accompanied by the non-analytic singularity of the PGP, by decomposing the PGP into each quasiparticle mode. The connection between the non-analytic behavior of the PGP at the critical time and the DQPT, regardless of whether the winding number is topological, can be explained by the fact that they both arise when the Loschmidt amplitude vanishes., Comment: 14 pages, 8 figures

Atmospheric powered descent guidance can be solved by successive convexification; however, its onboard application is impeded by the sharp increase in computation caused by nonlinear aerodynamic forces. The problem has to be converted into a sequence of convex subproblems instead of a single convex problem when aerodynamic forces are ignored. Besides, each subproblem is significantly more complicated, which increases computation. A fast real-time interior point method was presented to solve the correlated convex subproblems onboard in the work. The main contributions are as follows: Firstly, an algorithm was proposed to accelerate the solution of linear systems that cost most of the computation in each iterative step by exploiting the specific problem structure. Secondly, a warm-starting scheme was introduced to refine the initial value of a subproblem with a rough approximate solution of the former subproblem, which lessened the iterative steps required for each subproblem. The method proposed reduced the run time by a factor of 9 compared with the fastest publicly available solver tested in Monte Carlo simulations to evaluate the efficiency of solvers. Runtimes on the order of 0.6 s are achieved on a radiation-hardened flight processor, which demonstrated the potential of the real-time onboard application., Comment: The paper is accepted by IEEE Transactions on Aerospace and Electronic Systems, 2023

Hybrid modeling combining data-driven techniques and numerical methods is an emerging and promising research direction for efficient climate simulation. However, previous works lack practical platforms, making developing hybrid modeling a challenging programming problem. Furthermore, the lack of standard data sets and evaluation metrics may hamper researchers from comprehensively comparing various algorithms under a uniform condition. To address these problems, we propose a framework called NeuroClim for hybrid modeling under the real-world scenario, a basic setting to simulate the real climate that we live in. NeuroClim consists of three parts: (1) Platform. We develop a user-friendly platform NeuroGCM for efficiently developing hybrid modeling in climate simulation. (2) Dataset. We provide an open-source dataset for data-driven methods in hybrid modeling. We investigate the characteristics of the data, i.e., heterogeneity and stiffness, which reveals the difficulty of regressing climate simulation data; (3) Metrics. We propose a methodology for quantitatively evaluating hybrid modeling, including the approximation ability of machine learning models and the stability during simulation. We believe that NeuroClim allows researchers to work without high level of climate-related expertise and focus only on machine learning algorithm design, which will accelerate hybrid modeling research in the AI-Climate intersection. The codes and data are released at https://github.com/x-w19/NeuroClim., Comment: Work in progress

This paper studies the worst case iteration complexity of an infeasible interior point method (IPM) for seconder order cone programming (SOCP), which is more convenient for warmstarting compared with feasible IPMs. The method studied bases on the homogeneous and self-dual model and the Monteiro-Zhang family of searching directions. Its worst case iteration complexity is $O\left(k^{1/2}\log\left(\epsilon^{-1}\right)\right)$, to reduce the primal residual, dual residual, and complementarity gap by a factor of $\epsilon$, where $k$ is the number of cone constraints. The result is the same as the best known result for feasible IPMs. The condition under which warmstarting improves the complexity bound is also studied.

High-performance classical simulator for quantum circuits, in particular the tensor network contraction algorithm, has become an important tool for the validation of noisy quantum computing. In order to address the memory limitations, the slicing technique is used to reduce the tensor dimensions, but it could also lead to additional computation overhead that greatly slows down the overall performance. This paper proposes novel lifetime-based methods to reduce the slicing overhead and improve the computing efficiency, including an interpretation method to deal with slicing overhead, an in-place slicing strategy to find the smallest slicing set and an adaptive tensor network contraction path refiner customized for Sunway architecture. Experiments show that in most cases the slicing overhead with our in-place slicing strategy would be less than the cotengra, which is the most used graph path optimization software at present. Finally, the resulting simulation time is reduced to 96.1s for the Sycamore quantum processor RQC, with a sustainable single-precision performance of 308.6Pflops using over 41M cores to generate 1M correlated samples, which is more than 5 times performance improvement compared to 60.4 Pflops in 2021 Gordon Bell Prize work., Comment: 12 pages, 13 figures

Risk assessment is a substantial problem for financial institutions that has been extensively studied both for its methodological richness and its various practical applications. With the expansion of inclusive finance, recent attentions are paid to micro and small-sized enterprises (MSEs). Compared with large companies, MSEs present a higher exposure rate to default owing to their insecure financial stability. Conventional efforts learn classifiers from historical data with elaborate feature engineering. However, the main obstacle for MSEs involves severe deficiency in credit-related information, which may degrade the performance of prediction. Besides, financial activities have diverse explicit and implicit relations, which have not been fully exploited for risk judgement in commercial banks. In particular, the observations on real data show that various relationships between company users have additional power in financial risk analysis. In this paper, we consider a graph of banking data, and propose a novel HIDAM model for the purpose. Specifically, we attempt to incorporate heterogeneous information network with rich attributes on multi-typed nodes and links for modeling the scenario of business banking service. To enhance feature representation of MSEs, we extract interactive information through meta-paths and fully exploit path information. Furthermore, we devise a hierarchical attention mechanism respectively to learn the importance of contents inside each meta-path and the importance of different metapahs. Experimental results verify that HIDAM outperforms state-of-the-art competitors on real-world banking data., Comment: Corrected typos

The tantalizing promise of quantum computational speedup in solving certain problems has been strongly supported by recent experimental evidence from a high-fidelity 53-qubit superconducting processor1 and Gaussian boson sampling (GBS) with up to 76 detected photons. Analogous to the increasingly sophisticated Bell tests that continued to refute local hidden variable theories, quantum computational advantage tests are expected to provide increasingly compelling experimental evidence against the Extended Church-Turing thesis. In this direction, continued competition between upgraded quantum hardware and improved classical simulations is required. Here, we report a new GBS experiment that produces up to 113 detection events out of a 144-mode photonic circuit. We develop a new high-brightness and scalable quantum light source, exploring the idea of stimulated squeezed photons, which has simultaneously near-unity purity and efficiency. This GBS is programmable by tuning the phase of the input squeezed states. We demonstrate a new method to efficiently validate the samples by inferring from computationally friendly subsystems, which rules out hypotheses including distinguishable photons and thermal states. We show that our noisy GBS experiment passes the nonclassicality test using an inequality, and we reveal non-trivial genuine high-order correlation in the GBS samples, which are evidence of robustness against possible classical simulation schemes. The photonic quantum computer, Jiuzhang 2.0, yields a Hilbert space dimension up to $10^{43}$, and a sampling rate $10^{24}$ faster than using brute-force simulation on supercomputers., Comment: 23 pages, 6 figures. Comments are welcome

Gaussian boson sampling exploits squeezed states to provide a highly efficient way to demonstrate quantum computational advantage. We perform experiments with 50 input single-mode squeezed states with high indistinguishability and squeezing parameters, which are fed into a 100-mode ultralow-loss interferometer with full connectivity and random transformation, and sampled using 100 high-efficiency single-photon detectors. The whole optical set-up is phase-locked to maintain a high coherence between the superposition of all photon number states. We observe up to 76 output photon-clicks, which yield an output state space dimension of $10^{30}$ and a sampling rate that is $10^{14}$ faster than using the state-of-the-art simulation strategy and supercomputers. The obtained samples are validated against various hypotheses including using thermal states, distinguishable photons, and uniform distribution., Comment: 23 pages, 5 figures, supplemental information not included but a link is provided. This work is dedicated to the people in the fight against the COVID-19 outbreak during which the final stage of this experiment was carried out

Boson sampling is expected to be one of an important milestones that will demonstrate quantum supremacy. The present work establishes the benchmarking of Gaussian boson sampling (GBS) with threshold detection based on the Sunway TaihuLight supercomputer. To achieve the best performance and provide a competitive scenario for future quantum computing studies, the selected simulation algorithm is fully optimized based on a set of innovative approaches, including a parallel scheme and instruction-level optimizing method. Furthermore, data precision and instruction scheduling are handled in a sophisticated manner by an adaptive precision optimization scheme and a DAG-based heuristic search algorithm, respectively. Based on these methods, a highly efficient and parallel quantum sampling algorithm is designed. The largest run enables us to obtain one Torontonian function of a 100 x 100 submatrix from 50-photon GBS within 20 hours in 128-bit precision and 2 days in 256-bit precision.

Meta-reinforcement learning (meta-RL) aims to learn from multiple training tasks the ability to adapt efficiently to unseen test tasks. Despite the success, existing meta-RL algorithms are known to be sensitive to the task distribution shift. When the test task distribution is different from the training task distribution, the performance may degrade significantly. To address this issue, this paper proposes Model-based Adversarial Meta-Reinforcement Learning (AdMRL), where we aim to minimize the worst-case sub-optimality gap -- the difference between the optimal return and the return that the algorithm achieves after adaptation -- across all tasks in a family of tasks, with a model-based approach. We propose a minimax objective and optimize it by alternating between learning the dynamics model on a fixed task and finding the adversarial task for the current model -- the task for which the policy induced by the model is maximally suboptimal. Assuming the family of tasks is parameterized, we derive a formula for the gradient of the suboptimality with respect to the task parameters via the implicit function theorem, and show how the gradient estimator can be efficiently implemented by the conjugate gradient method and a novel use of the REINFORCE estimator. We evaluate our approach on several continuous control benchmarks and demonstrate its efficacy in the worst-case performance over all tasks, the generalization power to out-of-distribution tasks, and in training and test time sample efficiency, over existing state-of-the-art meta-RL algorithms., Comment: Accepted by NeurIPS 2020. Code at https://github.com/LinZichuan/AdMRL

The difficulty of deploying various deep learning (DL) models on diverse DL hardware has boosted the research and development of DL compilers in the community. Several DL compilers have been proposed from both industry and academia such as Tensorflow XLA and TVM. Similarly, the DL compilers take the DL models described in different DL frameworks as input, and then generate optimized codes for diverse DL hardware as output. However, none of the existing survey has analyzed the unique design architecture of the DL compilers comprehensively. In this paper, we perform a comprehensive survey of existing DL compilers by dissecting the commonly adopted design in details, with emphasis on the DL oriented multi-level IRs, and frontend/backend optimizations. Specifically, we provide a comprehensive comparison among existing DL compilers from various aspects. In addition, we present detailed analysis on the design of multi-level IRs and illustrate the commonly adopted optimization techniques. Finally, several insights are highlighted as the potential research directions of DL compiler. This is the first survey paper focusing on the design architecture of DL compilers, which we hope can pave the road for future research towards DL compiler.

Many reinforcement learning (RL) tasks have specific properties that can be leveraged to modify existing RL algorithms to adapt to those tasks and further improve performance, and a general class of such properties is the multiple reward channel. In those environments the full reward can be decomposed into sub-rewards obtained from different channels. Existing work on reward decomposition either requires prior knowledge of the environment to decompose the full reward, or decomposes reward without prior knowledge but with degraded performance. In this paper, we propose Distributional Reward Decomposition for Reinforcement Learning (DRDRL), a novel reward decomposition algorithm which captures the multiple reward channel structure under distributional setting. Empirically, our method captures the multi-channel structure and discovers meaningful reward decomposition, without any requirements on prior knowledge. Consequently, our agent achieves better performance than existing methods on environments with multiple reward channels., Comment: NeurlPS 2019

In this paper, we report the implementation and measured performance of our extreme-scale global simulation code on Sunway TaihuLight and two PEZY-SC2 systems: Shoubu System B and Gyoukou. The numerical algorithm is the parallel Barnes-Hut tree algorithm, which has been used in many large-scale astrophysical particle-based simulations. Our implementation is based on our FDPS framework. However, the extremely large numbers of cores of the systems used (10M on TaihuLight and 16M on Gyoukou) and their relatively poor memory and network bandwidth pose new challenges. We describe the new algorithms introduced to achieve high efficiency on machines with low memory bandwidth. The measured performance is 47.9, 10.6 PF, and 1.01PF on TaihuLight, Gyoukou and Shoubu System B (efficiency 40\%, 23.5\% and 35.5\%). The current code is developed for the simulation of planetary rings, but most of the new algorithms are useful for other simulations, and are now available in the FDPS framework.

We present the remote stochastic gradient (RSG) method, which computes the gradients at configurable remote observation points, in order to improve the convergence rate and suppress gradient noise at the same time for different curvatures. RSG is further combined with adaptive methods to construct ARSG for acceleration. The method is efficient in computation and memory, and is straightforward to implement. We analyze the convergence properties by modeling the training process as a dynamic system, which provides a guideline to select the configurable observation factor without grid search. ARSG yields $O(1/\sqrt{T})$ convergence rate in non-convex settings, that can be further improved to $O(\log(T)/T)$ in strongly convex settings. Numerical experiments demonstrate that ARSG achieves both faster convergence and better generalization, compared with popular adaptive methods, such as ADAM, NADAM, AMSGRAD, and RANGER for the tested problems. In particular, for training ResNet-50 on ImageNet, ARSG outperforms ADAM in convergence speed and meanwhile it surpasses SGD in generalization., Comment: The generalization is improved by modifying the preconditioner. For training ResNet-50 on ImageNet, ARSG outperforms ADAM in convergence speed and meanwhile it surpasses SGD in generalization. We also present a convergence bound in non-convex settings

The flourish of deep learning frameworks and hardware platforms has been demanding an efficient compiler that can shield the diversity in both software and hardware in order to provide application portability. Among the existing deep learning compilers, TVM is well known for its efficiency in code generation and optimization across diverse hardware devices. In the meanwhile, the Sunway many-core processor renders itself as a competitive candidate for its attractive computational power in both scientific computing and deep learning workloads. This paper combines the trends in these two directions. Specifically, we propose swTVM that extends the original TVM to support ahead-of-time compilation for architecture requiring cross-compilation such as Sunway. In addition, we leverage the architecture features during the compilation such as core group for massive parallelism, DMA for high bandwidth memory transfer and local device memory for data locality, in order to generate efficient codes for deep learning workloads on Sunway. The experiment results show that the codes generated by swTVM achieves 1.79x on average compared to the state-of-the-art deep learning framework on Sunway, across six representative benchmarks. This work is the first attempt from the compiler perspective to bridge the gap of deep learning and Sunway processor particularly with productivity and efficiency in mind. We believe this work will encourage more people to embrace the power of deep learning and Sunway many-core processor.

This paper reports our efforts on swCaffe, a highly efficient parallel framework for accelerating deep neural networks (DNNs) training on Sunway TaihuLight, the current fastest supercomputer in the world that adopts a unique many-core heterogeneous architecture, with 40,960 SW26010 processors connected through a customized communication network. First, we point out some insightful principles to fully exploit the performance of the innovative many-core architecture. Second, we propose a set of optimization strategies for redesigning a variety of neural network layers based on Caffe. Third, we put forward a topology-aware parameter synchronization scheme to scale the synchronous Stochastic Gradient Descent (SGD) method to multiple processors efficiently. We evaluate our framework by training a variety of widely used neural networks with the ImageNet dataset. On a single node, swCaffe can achieve 23\%\~{}119\% overall performance compared with Caffe running on K40m GPU. As compared with the Caffe on CPU, swCaffe runs 3.04\~{}7.84x faster on all the networks. Finally, we present the scalability of swCaffe for the training of ResNet-50 and AlexNet on the scale of 1024 nodes., Comment: 10 pages

We show that low-depth random quantum circuits can be efficiently simulated by a quantum teleportation-inspired algorithm. By using logical qubits to redirect and teleport the quantum information in quantum circuits, the original circuits can be renormalized to new circuits with a smaller number of logical qubits. We demonstrate the algorithm to simulate several random quantum circuits, including 1D-chain 1000-qubit 42-depth, 2D-grid 125*8-qubit 42-depth and 2D-Bristlecone 72-qubit 32-depth circuits. Our results present a memory-efficient method with a clear physical picture to simulate low-depth random quantum circuits., Comment: 9 pages, 11 figures

Data parallelism has become a dominant method to scale Deep Neural Network (DNN) training across multiple nodes. Since synchronizing a large number of gradients of the local model can be a bottleneck for large-scale distributed training, compressing communication data has gained widespread attention recently. Among several recent proposed compression algorithms, Residual Gradient Compression (RGC) is one of the most successful approaches---it can significantly compress the transmitting message size (0.1\% of the gradient size) of each node and still achieve correct accuracy and the same convergence speed. However, the literature on compressing deep networks focuses almost exclusively on achieving good theoretical compression rate, while the efficiency of RGC in real distributed implementation has been less investigated. In this paper, we develop an RGC-based system that is able to reduce the end-to-end training time on real-world multi-GPU systems. Our proposed design called RedSync, which introduces a set of optimizations to reduce communication bandwidth requirement while introducing limited overhead. We evaluate the performance of RedSync on two different multiple GPU platforms, including 128 GPUs of a supercomputer and an 8-GPU server. Our test cases include image classification tasks on Cifar10 and ImageNet, and language modeling tasks on Penn Treebank and Wiki2 datasets. For DNNs featured with high communication to computation ratio, which have long been considered with poor scalability, RedSync brings significant performance improvements., Comment: 10 pages. Journal of Parallel and Distributed Computing, 2019

Reinforcement learning (RL) algorithms have made huge progress in recent years by leveraging the power of deep neural networks (DNN). Despite the success, deep RL algorithms are known to be sample inefficient, often requiring many rounds of interaction with the environments to obtain satisfactory performance. Recently, episodic memory based RL has attracted attention due to its ability to latch on good actions quickly. In this paper, we present a simple yet effective biologically inspired RL algorithm called Episodic Memory Deep Q-Networks (EMDQN), which leverages episodic memory to supervise an agent during training. Experiments show that our proposed method can lead to better sample efficiency and is more likely to find good policies. It only requires 1/5 of the interactions of DQN to achieve many state-of-the-art performances on Atari games, significantly outperforming regular DQN and other episodic memory based RL algorithms., Comment: Accepted by IJCAI 2018

With the rapid progress made by industry and academia, quantum computers with dozens of qubits or even larger size are being realized. However, the fidelity of existing quantum computers often sharply decreases as the circuit depth increases. Thus, an ideal quantum circuit simulator on classical computers, especially on high-performance computers, is needed for benchmarking and validation. We design a large-scale simulator of universal random quantum circuits, often called 'quantum supremacy circuits', and implement it on Sunway TaihuLight. The simulator can be used to accomplish the following two tasks: 1) Computing a complete output state-vector; 2) Calculating one or a few amplitudes. We target the simulation of 49-qubit circuits. For task 1), we successfully simulate such a circuit of depth 39, and for task 2) we reach the 55-depth level. To the best of our knowledge, both of the simulation results reach the largest depth for 49-qubit quantum supremacy circuits., Comment: slightly modify the organisation

Recent deep learning (DL) models have moved beyond static network architectures to dynamic ones, handling data where the network structure changes every example, such as sequences of variable lengths, trees, and graphs. Existing dataflow-based programming models for DL---both static and dynamic declaration---either cannot readily express these dynamic models, or are inefficient due to repeated dataflow graph construction and processing, and difficulties in batched execution. We present Cavs, a vertex-centric programming interface and optimized system implementation for dynamic DL models. Cavs represents dynamic network structure as a static vertex function $\mathcal{F}$ and a dynamic instance-specific graph $\mathcal{G}$, and performs backpropagation by scheduling the execution of $\mathcal{F}$ following the dependencies in $\mathcal{G}$. Cavs bypasses expensive graph construction and preprocessing overhead, allows for the use of static graph optimization techniques on pre-defined operations in $\mathcal{F}$, and naturally exposes batched execution opportunities over different graphs. Experiments comparing Cavs to two state-of-the-art frameworks for dynamic NNs (TensorFlow Fold and DyNet) demonstrate the efficacy of this approach: Cavs achieves a near one order of magnitude speedup on training of various dynamic NN architectures, and ablations demonstrate the contribution of our proposed batching and memory management strategies., Comment: Short versions of this paper were presented at AISys workshop@SOSP 2017 and MLSys workshop@NIPS 2017

Fusion energy will be the ultimate clean energy source for mankind. One of the most visible concerns of the future fusion device is the threat of deleterious runaway electrons (REs) produced during unexpected disruptions of the fusion plasma. Both efficient long-term algorithms and super-large scale computing power are necessary to reveal the complex dynamics of REs in a realistic fusion reactor. In the present study, we deploy the world's fastest supercomputer, Sunway TaihuLight, and the newly developed relativistic volume-preserving algorithm to carry out long-term particle simulations of 10^7 sampled REs in 6D phase space, which involves simulation scale of 10^18 particle-steps, the largest ever achieved in fusion research. Our simulations show that in a realistic fusion reactor, the concern of REs is not as serious as previously thought. Specifically, REs are confined much better than previously predicted and the maximum average energy is in the range of 150MeV, less than half of previous estimate.