Search

Given the popularity of generative AI, Large Language Models (LLMs) often consume hundreds or thousands of GPUs for parallelizing and accelerating the training process. Communication overhead becomes more pronounced when training LLMs at scale. To eliminate communication overhead in distributed LLM training, we propose Domino, which provides a generic scheme to hide communication behind computation. By breaking data dependency of a single batch training into smaller independent pieces, Domino pipelines these independent pieces training and provides generic strategy of fine-grained communication and computation overlapping. Extensive results show that, comparing with Megatron-LM, Domino achieves up to 1.3x speedup for LLM training on Nvidia DGX-H100 GPUs., Comment: 12 pages

Model checkpoints are critical Deep Learning (DL) artifacts that enable fault tolerance for training and downstream applications, such as inference. However, writing checkpoints to persistent storage, and other I/O aspects of DL training, are mostly ignored by compute-focused optimization efforts for faster training of rapidly growing models and datasets. Towards addressing this imbalance, we propose FastPersist to accelerate checkpoint creation in DL training. FastPersist combines three novel techniques: (i) NVMe optimizations for faster checkpoint writes to SSDs, (ii) efficient write parallelism using the available SSDs in training environments, and (iii) overlapping checkpointing with independent training computations. Our evaluation using real world dense and sparse DL models shows that FastPersist creates checkpoints in persistent storage up to 116x faster than baseline, and enables per-iteration checkpointing with negligible overhead., Comment: 11 pages

Magnetic resonance imaging (MRI) has significantly benefited from the resurgence of artificial intelligence (AI). By leveraging AI's capabilities in large-scale optimization and pattern recognition, innovative methods are transforming the MRI acquisition workflow, including planning, sequence design, and correction of acquisition artifacts. These emerging algorithms demonstrate substantial potential in enhancing the efficiency and throughput of acquisition steps. This review discusses several pivotal AI-based methods in neuro MRI acquisition, focusing on their technological advances, impact on clinical practice, and potential risks., Comment: Magn Reson Mater Phy (2024)

Quantitative MRI enables direct quantification of contrast agent concentrations in contrast-enhanced scans. However, the lengthy scan times required by conventional methods are inadequate for tracking contrast agent transport dynamically in mouse brain. We developed a 3D MR fingerprinting (MRF) method for simultaneous T1 and T2 mapping across the whole mouse brain with 4.3-min temporal resolution. We designed a 3D MRF sequence with variable acquisition segment lengths and magnetization preparations on a 9.4T preclinical MRI scanner. Model-based reconstruction approaches were employed to improve the accuracy and speed of MRF acquisition. The method's accuracy for T1 and T2 measurements was validated in vitro, while its repeatability of T1 and T2 measurements was evaluated in vivo (n=3). The utility of the 3D MRF sequence for dynamic tracking of intracisternally infused Gd-DTPA in the whole mouse brain was demonstrated (n=5). Phantom studies confirmed accurate T1 and T2 measurements by 3D MRF with an undersampling factor up to 48. Dynamic contrast-enhanced (DCE) MRF scans achieved a spatial resolution of 192 x 192 x 500 um3 and a temporal resolution of 4.3 min, allowing for the analysis and comparison of dynamic changes in concentration and transport kinetics of intracisternally infused Gd-DTPA across brain regions. The sequence also enabled highly repeatable, high-resolution T1 and T2 mapping of the whole mouse brain (192 x 192 x 250 um3) in 30 min. We present the first dynamic and multi-parametric approach for quantitatively tracking contrast agent transport in the mouse brain using 3D MRF.

We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench), despite being small enough to be deployed on a phone. Our training dataset is a scaled-up version of the one used for phi-2, composed of heavily filtered publicly available web data and synthetic data. The model is also further aligned for robustness, safety, and chat format. We also provide parameter-scaling results with a 7B, 14B models trained for 4.8T tokens, called phi-3-small, phi-3-medium, both significantly more capable than phi-3-mini (e.g., respectively 75%, 78% on MMLU, and 8.7, 8.9 on MT-bench). To enhance multilingual, multimodal, and long-context capabilities, we introduce three models in the phi-3.5 series: phi-3.5-mini, phi-3.5-MoE, and phi-3.5-Vision. The phi-3.5-MoE, a 16 x 3.8B MoE model with 6.6 billion active parameters, achieves superior performance in language reasoning, math, and code tasks compared to other open-source models of similar scale, such as Llama 3.1 and the Mixtral series, and on par with Gemini-1.5-Flash and GPT-4o-mini. Meanwhile, phi-3.5-Vision, a 4.2 billion parameter model derived from phi-3.5-mini, excels in reasoning tasks and is adept at handling both single-image and text prompts, as well as multi-image and text prompts., Comment: 24 pages

The sixth-generation (6G) network is expected to achieve global coverage based on the space-air-ground integrated network, and the latest satellite network will play an important role in it. The introduction of inter-satellite links (ISLs) can significantly improve the throughput of the satellite network, and recently gets lots of attention from both academia and industry. In this paper, we illustrate the advantages of using the laser for ISLs due to its longer communication distance, higher data speed, and stronger security. Specifically, space-borne laser terminals with the acquisition, pointing and tracking mechanism which realize long-distance communication are illustrated, advanced modulation and multiplexing modes that make high communication rates possible are introduced, and the security of ISLs ensured by the characteristics of both laser and the optical channel is also analyzed. Moreover, some open issues such as advanced optical beam steering, routing and scheduling algorithm, and integrated sensing and communication are discussed to direct future research.

Zero Redundancy Optimizer (ZeRO) has been used to train a wide range of large language models on massive GPUs clusters due to its ease of use, efficiency, and good scalability. However, when training on low-bandwidth clusters, or at scale which forces batch size per GPU to be small, ZeRO's effective throughput is limited because of high communication volume from gathering weights in forward pass, backward pass, and averaging gradients. This paper introduces three communication volume reduction techniques, which we collectively refer to as ZeRO++, targeting each of the communication collectives in ZeRO. First is block-quantization based all-gather. Second is data remapping that trades-off communication for more memory. Third is a novel all-to-all based quantized gradient averaging paradigm as replacement of reduce-scatter collective, which preserves accuracy despite communicating low precision data. Collectively, ZeRO++ reduces communication volume of ZeRO by 4x, enabling up to 2.16x better throughput at 384 GPU scale., Comment: 12 pages

We study the excludable public project model where the decision is binary (build or not build). In a classic excludable and binary public project model, an agent either consumes the project in its whole or is completely excluded. We study a setting where the mechanism can set different project release time for different agents, in the sense that high-paying agents can consume the project earlier than low-paying agents. The release delay, while hurting the social welfare, is implemented to incentivize payments to cover the project cost. The mechanism design objective is to minimize the maximum release delay and the total release delay among all agents. We first consider the setting where we know the prior distribution of the agents' types. Our objectives are minimizing the expected maximum release delay and the expected total release delay. We propose the single deadline mechanisms. We show that the optimal single deadline mechanism is asymptotically optimal for both objectives, regardless of the prior distribution. For small number of agents, we propose the sequential unanimous mechanisms by extending the largest unanimous mechanisms from [Ohseto 2000]. We propose an automated mechanism design approach via evolutionary computation to optimize within the sequential unanimous mechanisms. We next study prior-free mechanism design. We propose the group-based optimal deadline mechanism and show that it is competitive against an undominated mechanism under minor technical assumptions., Comment: arXiv admin note: substantial text overlap with arXiv:2204.07315

Photolithography assisted chemo-mechanical etching (PLACE), a technique specifically developed for fabricating highquality large-scale photonic integrated circuits (PICs) on thin-film lithium niobate (TFLN), has enabled fabrication of a series of building blocks of PICs ranging from high-quality (high-Q) microresonators and low-loss waveguides to electrooptically (EO) tunable lasers and waveguide amplifiers. Aiming at high-throughput manufacturing of the PIC devices and systems, we have developed an ultra-high-speed high-resolution laser lithography fabrication system employing a high repetition rate femtosecond laser and a high-speed polygon laser scanner, by which a lithography fabrication efficiency of 4.8 cm2/h has been achieved at a spatial resolution of 200 nm. We demonstrate wafer-scale fabrication of TFLN-based photonic structures, optical phase masks as well as color printing

Adaptive or dynamic signal sampling in sensing systems can adapt subsequent sampling strategies based on acquired signals, thereby potentially improving image quality and speed. This paper proposes a Bayesian method for adaptive sampling based on greedy variance reduction and stochastic gradient Langevin dynamics (SGLD). The image priors involved can be either analytical or neural network-based. Notably, the learned image priors generalize well to out-of-distribution test cases that have different statistics than the training dataset. As a real-world validation, the method is applied to accelerate the acquisition of magnetic resonance imaging (MRI). Compared to non-adaptive sampling, the proposed method effectively improved the image quality by 2-3 dB in PSNR, and improved the restoration of subtle details., Comment: 5 pages, 4 figures

Optimizing 3D k-space sampling trajectories for efficient MRI is important yet challenging. This work proposes a generalized framework for optimizing 3D non-Cartesian sampling patterns via data-driven optimization. We built a differentiable MRI system model to enable gradient-based methods for sampling trajectory optimization. By combining training losses, the algorithm can simultaneously optimize multiple properties of sampling patterns, including image quality, hardware constraints (maximum slew rate and gradient strength), reduced peripheral nerve stimulation (PNS), and parameter-weighted contrast. The proposed method can either optimize the gradient waveform (spline-based freeform optimization) or optimize properties of given sampling trajectories (such as the rotation angle of radial trajectories). Notably, the method optimizes sampling trajectories synergistically with either model-based or learning-based reconstruction methods. We proposed several strategies to alleviate the severe non-convexity and huge computation demand posed by the high-dimensional optimization. The corresponding code is organized as an open-source, easy-to-use toolbox. We applied the optimized trajectory to multiple applications including structural and functional imaging. In the simulation studies, the reconstruction PSNR of a 3D kooshball trajectory was increased by 4 dB with SNOPY optimization. In the prospective studies, by optimizing the rotation angles of a stack-of-stars (SOS) trajectory, SNOPY improved the PSNR by 1.4dB compared to the best empirical method. Optimizing the gradient waveform of a rotational EPI trajectory improved subjects' rating of the PNS effect from 'strong' to 'mild.' In short, SNOPY provides an efficient data-driven and optimization-based method to tailor non-Cartesian sampling trajectories., Comment: 13 pages, 8 figures

Mechanism design is a central research branch in microeconomics. An effective mechanism can significantly improve performance and efficiency of social decisions under desired objectives, such as to maximize social welfare or to maximize revenue for agents. However, mechanism design is challenging for many common models including the public project problem model which we study in this thesis. A typical public project problem is a group of agents crowdfunding a public project (e.g., building a bridge). The mechanism will decide the payment and allocation for each agent (e.g., how much the agent pays, and whether the agent can use it) according to their valuations. The mechanism can be applied to various economic scenarios, including those related to cyber security. There are different constraints and optimized objectives for different public project scenarios (sub-problems), making it unrealistic to design a universal mechanism that fits all scenarios, and designing mechanisms for different settings manually is a taxing job. Therefore, we explore automated mechanism design (AMD) of public project problems under different constraints. In this thesis, we focus on the public project problem, which includes many sub-problems (excludable/non-excludable, divisible/indivisible, binary/non-binary). We study the classical public project model and extend this model to other related areas such as the zero-day exploit markets. For different sub-problems of the public project problem, we adopt different novel machine learning techniques to design optimal or near-optimal mechanisms via automated mechanism design. We evaluate our mechanisms by theoretical analysis or experimentally comparing our mechanisms against existing mechanisms. The experiments and theoretical results show that our mechanisms are better than state-of-the-art automated or manual mechanisms., Comment: Phd Thesis

Many important problems in multiagent systems involve resource allocations. Self-interested agents may lie about their valuations if doing so increases their own utilities. Therefore, it is necessary to design mechanisms (collective decision-making rules) with desired properties and objectives. The VCG redistribution mechanisms are efficient (the agents who value the resources the most will be allocated), strategy-proof (the agents have no incentives to lie about their valuations), and weakly budget-balanced (no deficits). We focus on the VCG redistribution mechanisms for the classic public project problem, where a group of agents needs to decide whether or not to build a non-excludable public project. We design mechanisms via neural networks with two welfare-maximizing objectives: optimal in the worst case and optimal in expectation. Previous studies showed two worst-case optimal mechanisms for 3 agents, but worst-case optimal mechanisms have not been identified for more than 3 agents. For maximizing expected welfare, there are no existing results. We use neural networks to design VCG redistribution mechanisms. Neural networks have been used to design the redistribution mechanisms for multi-unit auctions with unit demand. We show that for the public project problem, the previously proposed neural networks, which led to optimal/near-optimal mechanisms for multi-unit auctions with unit demand, perform abysmally for the public project problem. We significantly improve the existing networks on multiple fronts: We conduct a GAN network to generate worst-case type profiles and feed prior distribution into loss function to provide quality gradients for the optimal-in-expectation objective......

Model predictive control (MPC) is a powerful control method that handles dynamical systems with constraints. However, solving MPC iteratively in real time, i.e., implicit MPC, remains a computational challenge. To address this, common solutions include explicit MPC and function approximation. Both methods, whenever applicable, may improve the computational efficiency of the implicit MPC by several orders of magnitude. Nevertheless, explicit MPC often requires expensive pre-computation and does not easily apply to higher-dimensional problems. Meanwhile, function approximation, although scales better with dimension, still requires pre-training on a large dataset and generally cannot guarantee to find an accurate surrogate policy, the failure of which often leads to closed-loop instability. To address these issues, we propose a triple-mode hybrid control scheme, named Memory-Augmented MPC, by combining a linear quadratic regulator, a neural network, and an MPC. From its standard form, we further derive two variants of such hybrid control scheme: one customized for chaotic systems and the other for slow systems. The proposed scheme does not require pre-computation and can improve the amortized running time of the composed MPC with a well-trained neural network. In addition, the scheme maintains closed-loop stability with any neural networks of proper input and output dimensions, alleviating the need for certifying optimality of the neural network in safety-critical applications., Comment: 13 pages, 10 figures, 2 tables

There is growing interest in learning Fourier domain sampling strategies (particularly for magnetic resonance imaging, MRI) using optimization approaches. For non-Cartesian sampling patterns, the system models typically involve non-uniform FFT (NUFFT) operations. Commonly used NUFFT algorithms contain frequency domain interpolation, which is not differentiable with respect to the sampling pattern, complicating the use of gradient methods. This paper describes an efficient and accurate approach for computing approximate gradients involving NUFFTs. Multiple numerical experiments validated the improved accuracy and efficiency of the proposed approximation. As an application to computational imaging, the NUFFT Jacobians were used to optimize non-Cartesian MRI sampling trajectories via data-driven stochastic optimization. Specifically, the sampling patterns were learned with respect to various model-based reconstruction algorithms, including quadratic regularized reconstruction and compressed sensing-based reconstruction. The proposed approach enables sampling optimization for image sizes that are infeasible with standard auto-differentiation methods due to memory limits. The synergistic acquisition and reconstruction lead to remarkably improved image quality. In fact, we show that model-based image reconstruction (MBIR) methods with suitably optimized imaging parameters can perform nearly as well as CNN-based methods., Comment: 12 pages, 5 figures