We conduct a systematic study of solving the learning parity with noise problem (LPN) using neural networks. Our main contribution is designing families of two-layer neural networks that practically outperform classical algorithms in high-noise, low-dimension regimes. We consider three settings where the numbers of LPN samples are abundant, very limited, and in between. In each setting we provide neural network models that solve LPN as fast as possible. For some settings we are also able to provide theories that explain the rationale of the design of our models. Comparing with the previous experiments of Esser, Kubler, and May (CRYPTO 2017), for dimension $n = 26$, noise rate $\tau = 0.498$, the ''Guess-then-Gaussian-elimination'' algorithm takes 3.12 days on 64 CPU cores, whereas our neural network algorithm takes 66 minutes on 8 GPUs. Our algorithm can also be plugged into the hybrid algorithms for solving middle or large dimension LPN instances.
Recent studies on 2D pose estimation have achieved excellent performance on public benchmarks, yet its application in the industrial community still suffers from heavy model parameters and high latency. In order to bridge this gap, we empirically explore key factors in pose estimation including paradigm, model architecture, training strategy, and deployment, and present a high-performance real-time multi-person pose estimation framework, RTMPose, based on MMPose. Our RTMPose-m achieves 75.8% AP on COCO with 90+ FPS on an Intel i7-11700 CPU and 430+ FPS on an NVIDIA GTX 1660 Ti GPU, and RTMPose-l achieves 67.0% AP on COCO-WholeBody with 130+ FPS. To further evaluate RTMPose's capability in critical real-time applications, we also report the performance after deploying on the mobile device. Our RTMPose-s achieves 72.2% AP on COCO with 70+ FPS on a Snapdragon 865 chip, outperforming existing open-source libraries. Code and models are released at https://github.com/open-mmlab/mmpose/tree/1.x/projects/rtmpose.
The high computational and memory requirements of large language model (LLM) inference make it feasible only with multiple high-end accelerators. Motivated by the emerging demand for latency-insensitive tasks with batched processing, this paper initiates the study of high-throughput LLM inference using limited resources, such as a single commodity GPU. We present FlexGen, a high-throughput generation engine for running LLMs with limited GPU memory. FlexGen can be flexibly configured under various hardware resource constraints by aggregating memory and computation from the GPU, CPU, and disk. By solving a linear programming problem, it searches for efficient patterns to store and access tensors. FlexGen further compresses the weights and the attention cache to 4 bits with negligible accuracy loss. These techniques enable FlexGen to have a larger space of batch size choices and thus significantly increase maximum throughput. As a result, when running OPT-175B on a single 16GB GPU, FlexGen achieves significantly higher throughput compared to state-of-the-art offloading systems, reaching a generation throughput of 1 token/s for the first time with an effective batch size of 144. On the HELM benchmark, FlexGen can benchmark a 30B model with a 16GB GPU on 7 representative sub-scenarios in 21 hours. The code is available at https://github.com/FMInference/FlexGen
Geologic carbon sequestration (GCS) is a safety-critical technology that aims to reduce the amount of carbon dioxide in the atmosphere, which also places high demands on reliability. Multiphase flow in porous media is essential to understand CO$_2$ migration and pressure fields in the subsurface associated with GCS. However, numerical simulation for such problems in 4D is computationally challenging and expensive, due to the multiphysics and multiscale nature of the highly nonlinear governing partial differential equations (PDEs). It prevents us from considering multiple subsurface scenarios and conducting real-time optimization. Here, we develop a Fourier-enhanced multiple-input neural operator (Fourier-MIONet) to learn the solution operator of the problem of multiphase flow in porous media. Fourier-MIONet utilizes the recently developed framework of the multiple-input deep neural operators (MIONet) and incorporates the Fourier neural operator (FNO) in the network architecture. Once Fourier-MIONet is trained, it can predict the evolution of saturation and pressure of the multiphase flow under various reservoir conditions, such as permeability and porosity heterogeneity, anisotropy, injection configurations, and multiphase flow properties. Compared to the enhanced FNO (U-FNO), the proposed Fourier-MIONet has 90% fewer unknown parameters, and it can be trained in significantly less time (about 3.5 times faster) with much lower CPU memory ($<$ 15%) and GPU memory ($<$ 35%) requirements, to achieve similar prediction accuracy. In addition to the lower computational cost, Fourier-MIONet can be trained with only 6 snapshots of time to predict the PDE solutions for 30 years. The excellent generalizability of Fourier-MIONet is enabled by its adherence to the physical principle that the solution to a PDE is continuous over time.
To design fast neural networks, many works have been focusing on reducing the number of floating-point operations (FLOPs). We observe that such reduction in FLOPs, however, does not necessarily lead to a similar level of reduction in latency. This mainly stems from inefficiently low floating-point operations per second (FLOPS). To achieve faster networks, we revisit popular operators and demonstrate that such low FLOPS is mainly due to frequent memory access of the operators, especially the depthwise convolution. We hence propose a novel partial convolution (PConv) that extracts spatial features more efficiently, by cutting down redundant computation and memory access simultaneously. Building upon our PConv, we further propose FasterNet, a new family of neural networks, which attains substantially higher running speed than others on a wide range of devices, without compromising on accuracy for various vision tasks. For example, on ImageNet-1k, our tiny FasterNet-T0 is $2.8\times$, $3.3\times$, and $2.4\times$ faster than MobileViT-XXS on GPU, CPU, and ARM processors, respectively, while being $2.9\%$ more accurate. Our large FasterNet-L achieves impressive $83.5\%$ top-1 accuracy, on par with the emerging Swin-B, while having $36\%$ higher inference throughput on GPU, as well as saving $37\%$ compute time on CPU. Code is available at \url{https://github.com/JierunChen/FasterNet}.
Weakly-supervised classification of histopathology slides is a computationally intensive task, with a typical whole slide image (WSI) containing billions of pixels to process. We propose Discriminative Region Active Sampling for Multiple Instance Learning (DRAS-MIL), a computationally efficient slide classification method using attention scores to focus sampling on highly discriminative regions. We apply this to the diagnosis of ovarian cancer histological subtypes, which is an essential part of the patient care pathway as different subtypes have different genetic and molecular profiles, treatment options, and patient outcomes. We use a dataset of 714 WSIs acquired from 147 epithelial ovarian cancer patients at Leeds Teaching Hospitals NHS Trust to distinguish the most common subtype, high-grade serous carcinoma, from the other four subtypes (low-grade serous, endometrioid, clear cell, and mucinous carcinomas) combined. We demonstrate that DRAS-MIL can achieve similar classification performance to exhaustive slide analysis, with a 3-fold cross-validated AUC of 0.8679 compared to 0.8781 with standard attention-based MIL classification. Our approach uses at most 18% as much memory as the standard approach, while taking 33% of the time when evaluating on a GPU and only 14% on a CPU alone. Reducing prediction time and memory requirements may benefit clinical deployment and the democratisation of AI, reducing the extent to which computational hardware limits end-user adoption.
Cost models predict the cost of executing given assembly code basic blocks on a specific microarchitecture. Recently, neural cost models have been shown to be fairly accurate and easy to construct. They can replace heavily engineered analytical cost models used in mainstream compiler workflows. However, their black-box nature discourages their adoption. In this work, we develop the first framework, COMET, for generating faithful, generalizable, and intuitive explanations for neural cost models. We generate and compare COMET's explanations for the popular neural cost model, Ithemal against those for an accurate CPU simulation-based cost model, uiCA. Our empirical findings show an inverse correlation between the prediction errors of Ithemal and uiCA and the granularity of basic block features in COMET's explanations for them, thus indicating potential reasons for the higher error of Ithemal with respect to uiCA.
The growing deployment of electric mobility calls for power system analyses to investigate to what extent the simultaneous charging of electric vehicles leads to degraded network operation and to validate the efficiency of countermeasures. To reduce complexity and CPU time, a common approach while performing these analyses consists in replacing electric vehicles and their charging stations with constant PQ loads. However, this approach is inaccurate, as the power absorbed by these elements actually depends not only on voltage but also on the state of charge, charging method, cathode chemistry of the battery pack, and converter controls in the electric vehicle and charging station. By considering all these aspects, this article develops a novel static load model and a vector fitting-based dynamic load model for electric vehicles connected to fast charging stations. These computationally efficient representations can replace the standard constant PQ load model of electric vehicles to assess more accurately their impact in static and dynamic grid studies. Simulation results of the IEEE14 system modified by adding fleets of electric vehicles prove the accuracy of the proposed models and highlight the shortcomings of the standard electric vehicle representation as a constant PQ load in some cases.
Challenging the Nvidia monopoly, dedicated AI-accelerator chips have begun emerging for tackling the computational challenge that the inference and, especially, the training of modern deep neural networks (DNNs) poses to modern computers. The field has been ridden with studies assessing the performance of these contestants across various DNN model types. However, AI-experts are aware of the limitations of current DNNs and have been working towards the fourth AI wave which will, arguably, rely on more biologically inspired models, predominantly on spiking neural networks (SNNs). At the same time, GPUs have been heavily used for simulating such models in the field of computational neuroscience, yet AI-chips have not been tested on such workloads. The current paper aims at filling this important gap by evaluating multiple, cutting-edge AI-chips (Graphcore IPU, GroqChip, Nvidia GPU with Tensor Cores and Google TPU) on simulating a highly biologically detailed model of a brain region, the inferior olive (IO). This IO application stress-tests the different AI-platforms for highlighting architectural tradeoffs by varying its compute density, memory requirements and floating-point numerical accuracy. Our performance analysis reveals that the simulation problem maps extremely well onto the GPU and TPU architectures, which for networks of 125,000 cells leads to a 28x respectively 1,208x speedup over CPU runtimes. At this speed, the TPU sets a new record for largest real-time IO simulation. The GroqChip outperforms both platforms for small networks but, due to implementing some floating-point operations at reduced accuracy, is found not yet usable for brain simulation.
Pitch is a foundational aspect of our perception of audio signals. Pitch contours are commonly used to analyze speech and music signals and as input features for many audio tasks, including music transcription, singing voice synthesis, and prosody editing. In this paper, we describe a set of techniques for improving the accuracy of widely-used neural pitch and periodicity estimators to achieve state-of-the-art performance on both speech and music. We also introduce a novel entropy-based method for extracting periodicity and per-frame voiced-unvoiced classifications from statistical inference-based pitch estimators (e.g., neural networks), and show how to train a neural pitch estimator to simultaneously handle both speech and music data (i.e., cross-domain estimation) without performance degradation. Our estimator implementations run 11.2x faster than real-time on a Intel i9-9820X 10-core 3.30 GHz CPU$\unicode{x2014}$approaching the speed of state-of-the-art DSP-based pitch estimators$\unicode{x2014}$or 408x faster than real-time on a NVIDIA GeForce RTX 3090 GPU. We release all of our code and models as Pitch-Estimating Neural Networks (penn), an open-source, pip-installable Python module for training, evaluating, and performing inference with pitch- and periodicity-estimating neural networks. The code for penn is available at https://github.com/interactiveaudiolab/penn.
On-device machine learning (ML) inference can enable the use of private user data on user devices without revealing them to remote servers. However, a pure on-device solution to private ML inference is impractical for many applications that rely on embedding tables that are too large to be stored on-device. In particular, recommendation models typically use multiple embedding tables each on the order of 1-10 GBs of data, making them impractical to store on-device. To overcome this barrier, we propose the use of private information retrieval (PIR) to efficiently and privately retrieve embeddings from servers without sharing any private information. As off-the-shelf PIR algorithms are usually too computationally intensive to directly use for latency-sensitive inference tasks, we 1) propose novel GPU-based acceleration of PIR, and 2) co-design PIR with the downstream ML application to obtain further speedup. Our GPU acceleration strategy improves system throughput by more than $20 \times$ over an optimized CPU PIR implementation, and our PIR-ML co-design provides an over $5 \times$ additional throughput improvement at fixed model quality. Together, for various on-device ML applications such as recommendation and language modeling, our system on a single V100 GPU can serve up to $100,000$ queries per second -- a $>100 \times$ throughput improvement over a CPU-based baseline -- while maintaining model accuracy.
Efficient detectors for edge devices are often optimized for parameters or speed count metrics, which remain in weak correlation with the energy of detectors. However, some vision applications of convolutional neural networks, such as always-on surveillance cameras, are critical for energy constraints. This paper aims to serve as a baseline by designing detectors to reach tradeoffs between energy and performance from two perspectives: 1) We extensively analyze various CNNs to identify low-energy architectures, including selecting activation functions, convolutions operators, and feature fusion structures on necks. These underappreciated details in past work seriously affect the energy consumption of detectors; 2) To break through the dilemmatic energy-performance problem, we propose a balanced detector driven by energy using discovered low-energy components named \textit{FemtoDet}. In addition to the novel construction, we improve FemtoDet by considering convolutions and training strategy optimizations. Specifically, we develop a new instance boundary enhancement (IBE) module for convolution optimization to overcome the contradiction between the limited capacity of CNNs and detection tasks in diverse spatial representations, and propose a recursive warm-restart (RecWR) for optimizing training strategy to escape the sub-optimization of light-weight detectors by considering the data shift produced in popular augmentations. As a result, FemtoDet with only 68.77k parameters achieves a competitive score of 46.3 AP50 on PASCAL VOC and 1.11 W $\&$ 64.47 FPS on Qualcomm Snapdragon 865 CPU platforms. Extensive experiments on COCO and TJU-DHD datasets indicate that the proposed method achieves competitive results in diverse scenes.
In this paper, we propose a novel LiDAR(-inertial) odometry and mapping framework to achieve the goal of simultaneous localization and meshing in real-time. This proposed framework termed ImMesh comprises four tightly-coupled modules: receiver, localization, meshing, and broadcaster. The localization module utilizes the prepossessed sensor data from the receiver, estimates the sensor pose online by registering LiDAR scans to maps, and dynamically grows the map. Then, our meshing module takes the registered LiDAR scan for incrementally reconstructing the triangle mesh on the fly. Finally, the real-time odometry, map, and mesh are published via our broadcaster. The key contribution of this work is the meshing module, which represents a scene by an efficient hierarchical voxels structure, performs fast finding of voxels observed by new scans, and reconstructs triangle facets in each voxel in an incremental manner. This voxel-wise meshing operation is delicately designed for the purpose of efficiency; it first performs a dimension reduction by projecting 3D points to a 2D local plane contained in the voxel, and then executes the meshing operation with pull, commit and push steps for incremental reconstruction of triangle facets. To the best of our knowledge, this is the first work in literature that can reconstruct online the triangle mesh of large-scale scenes, just relying on a standard CPU without GPU acceleration. To share our findings and make contributions to the community, we make our code publicly available on our GitHub: https://github.com/hku-mars/ImMesh.
Cybercriminals are moving towards zero-day attacks affecting resource-constrained devices such as single-board computers (SBC). Assuming that perfect security is unrealistic, Moving Target Defense (MTD) is a promising approach to mitigate attacks by dynamically altering target attack surfaces. Still, selecting suitable MTD techniques for zero-day attacks is an open challenge. Reinforcement Learning (RL) could be an effective approach to optimize the MTD selection through trial and error, but the literature fails when i) evaluating the performance of RL and MTD solutions in real-world scenarios, ii) studying whether behavioral fingerprinting is suitable for representing SBC's states, and iii) calculating the consumption of resources in SBC. To improve these limitations, the work at hand proposes an online RL-based framework to learn the correct MTD mechanisms mitigating heterogeneous zero-day attacks in SBC. The framework considers behavioral fingerprinting to represent SBCs' states and RL to learn MTD techniques that mitigate each malicious state. It has been deployed on a real IoT crowdsensing scenario with a Raspberry Pi acting as a spectrum sensor. More in detail, the Raspberry Pi has been infected with different samples of command and control malware, rootkits, and ransomware to later select between four existing MTD techniques. A set of experiments demonstrated the suitability of the framework to learn proper MTD techniques mitigating all attacks (except a harmfulness rootkit) while consuming <1 MB of storage and utilizing <55% CPU and <80% RAM.
Achieving resource efficiency while preserving end-user experience is non-trivial for cloud application operators. As cloud applications progressively adopt microservices, resource managers are faced with two distinct levels of system behavior: end-to-end application latency and per-service resource usage. Translating between the two levels, however, is challenging because user requests traverse heterogeneous services that collectively (but unevenly) contribute to the end-to-end latency. We present Autothrottle, a bi-level resource management framework for microservices with latency SLOs (service-level objectives). It architecturally decouples application SLO feedback from service resource control, and bridges them through the notion of performance targets. Specifically, an application-wide learning-based controller is employed to periodically set performance targets -- expressed as CPU throttle ratios -- for per-service heuristic controllers to attain. We evaluate Autothrottle on three microservice applications, with workload traces from production scenarios. Results show superior CPU savings, up to 26.21% over the best-performing baseline and up to 93.84% over all baselines.
Guided source separation (GSS) is a type of target-speaker extraction method that relies on pre-computed speaker activities and blind source separation to perform front-end enhancement of overlapped speech signals. It was first proposed during the CHiME-5 challenge and provided significant improvements over the delay-and-sum beamforming baseline. Despite its strengths, however, the method has seen limited adoption for meeting transcription benchmarks primarily due to its high computation time. In this paper, we describe our improved implementation of GSS that leverages the power of modern GPU-based pipelines, including batched processing of frequencies and segments, to provide 300x speed-up over CPU-based inference. The improved inference time allows us to perform detailed ablation studies over several parameters of the GSS algorithm -- such as context duration, number of channels, and noise class, to name a few. We provide end-to-end reproducible pipelines for speaker-attributed transcription of popular meeting benchmarks: LibriCSS, AMI, and AliMeeting. Our code and recipes are publicly available: https://github.com/desh2608/gss.
In recent years, large language models have achieved great success due to their unprecedented size. However, training these models poses a challenge for most researchers as it requires a substantial number of GPUs. To reduce GPU memory usage, memory partitioning, and memory offloading have been proposed. These approaches eliminate memory redundancies and offload memory usage to the CPU and NVMe memory, respectively, enabling training on small GPU clusters. However, directly deploying these solutions often leads to suboptimal efficiency. Only experienced experts can unleash the full potential of hardware by carefully tuning the distributed configuration. Thus, we present a novel solution, Elixir, which automates efficient large-model training based on pre-runtime model profiling. Elixir aims to identify the optimal combination of partitioning and offloading techniques to maximize training throughput. In our experiments, Elixir significantly outperforms the current state-of-the-art baseline. Our optimal configuration achieves up to a 3.4$\times$ speedup on GPT-2 models compared with SOTA solutions. We hope that our work will benefit individuals who lack computing resources and expertise, granting them access to large models. The beta version of Elixir is now available at https://github.com/hpcaitech/ColossalAI/tree/feature/elixir.
The deep learning revolution has greatly been accelerated by the 'hardware lottery': Recent advances in modern hardware accelerators and compilers paved the way for large-scale batch gradient optimization. Evolutionary optimization, on the other hand, has mainly relied on CPU-parallelism, e.g. using Dask scheduling and distributed multi-host infrastructure. Here we argue that also modern evolutionary computation can significantly benefit from the massive computational throughput provided by GPUs and TPUs. In order to better harness these resources and to enable the next generation of black-box optimization algorithms, we release evosax: A JAX-based library of evolution strategies which allows researchers to leverage powerful function transformations such as just-in-time compilation, automatic vectorization and hardware parallelization. evosax implements 30 evolutionary optimization algorithms including finite-difference-based, estimation-of-distribution evolution strategies and various genetic algorithms. Every single algorithm can directly be executed on hardware accelerators and automatically vectorized or parallelized across devices using a single line of code. It is designed in a modular fashion and allows for flexible usage via a simple ask-evaluate-tell API. We thereby hope to facilitate a new wave of scalable evolutionary optimization algorithms.
The rapidly evolving industry demands high accuracy of the models without the need for time-consuming and computationally expensive experiments required for fine-tuning. Moreover, a model and training pipeline, which was once carefully optimized for a specific dataset, rarely generalizes well to training on a different dataset. This makes it unrealistic to have carefully fine-tuned models for each use case. To solve this, we propose an alternative approach that also forms a backbone of Intel Geti platform: a dataset-agnostic template for object detection trainings, consisting of carefully chosen and pre-trained models together with a robust training pipeline for further training. Our solution works out-of-the-box and provides a strong baseline on a wide range of datasets. It can be used on its own or as a starting point for further fine-tuning for specific use cases when needed. We obtained dataset-agnostic templates by performing parallel training on a corpus of datasets and optimizing the choice of architectures and training tricks with respect to the average results on the whole corpora. We examined a number of architectures, taking into account the performance-accuracy trade-off. Consequently, we propose 3 finalists, VFNet, ATSS, and SSD, that can be deployed on CPU using the OpenVINO toolkit. The source code is available as a part of the OpenVINO Training Extensions (https://github.com/openvinotoolkit/training_extensions}
In this report, we present a fast and accurate object detection method dubbed DAMO-YOLO, which achieves higher performance than the state-of-the-art YOLO series. DAMO-YOLO is extended from YOLO with some new technologies, including Neural Architecture Search (NAS), efficient Reparameterized Generalized-FPN (RepGFPN), a lightweight head with AlignedOTA label assignment, and distillation enhancement. In particular, we use MAE-NAS, a method guided by the principle of maximum entropy, to search our detection backbone under the constraints of low latency and high performance, producing ResNet/CSP-like structures with spatial pyramid pooling and focus modules. In the design of necks and heads, we follow the rule of ``large neck, small head''.We import Generalized-FPN with accelerated queen-fusion to build the detector neck and upgrade its CSPNet with efficient layer aggregation networks (ELAN) and reparameterization. Then we investigate how detector head size affects detection performance and find that a heavy neck with only one task projection layer would yield better results.In addition, AlignedOTA is proposed to solve the misalignment problem in label assignment. And a distillation schema is introduced to improve performance to a higher level. Based on these new techs, we build a suite of models at various scales to meet the needs of different scenarios. For general industry requirements, we propose DAMO-YOLO-T/S/M/L. They can achieve 43.6/47.7/50.2/51.9 mAPs on COCO with the latency of 2.78/3.83/5.62/7.95 ms on T4 GPUs respectively. Additionally, for edge devices with limited computing power, we have also proposed DAMO-YOLO-Ns/Nm/Nl lightweight models. They can achieve 32.3/38.2/40.5 mAPs on COCO with the latency of 4.08/5.05/6.69 ms on X86-CPU. Our proposed general and lightweight models have outperformed other YOLO series models in their respective application scenarios.