This paper presents a novel neural network architecture for the purpose of pervasive visualisation of a 3D human upper limb musculoskeletal system model. Bringing simulation capabilities to resource-poor systems like mobile devices is of growing interest across many research fields, to widen applicability of methods and results. Until recently, this goal was thought to be out of reach for realistic continuum-mechanical simulations of musculoskeletal systems, due to prohibitive computational cost. Within this work we use a sparse grid surrogate to capture the surface deformation of the m.~biceps brachii in order to train a deep learning model, used for real-time visualisation of the same muscle. Both these surrogate models take 5 muscle activation levels as input and output Cartesian coordinate vectors for each mesh node on the muscle's surface. Thus, the neural network architecture features a significantly lower input than output dimension. 5 muscle activation levels were sufficient to achieve an average error of 0.97 +/- 0.16 mm, or 0.57 +/- 0.10 % for the 2809 mesh node positions of the biceps. The model achieved evaluation times of 9.88 ms per predicted deformation state on CPU only and 3.48 ms with GPU-support, leading to theoretical frame rates of 101 fps and 287 fps respectively. Deep learning surrogates thus provide a way to make continuum-mechanical simulations accessible for visual real-time applications.
Methane (CH4) is the second most critical greenhouse gas after carbon dioxide, contributing to 16-25% of the observed atmospheric warming. Wetlands are the primary natural source of methane emissions globally. However, wetland methane emission estimates from biogeochemistry models contain considerable uncertainty. One of the main sources of this uncertainty arises from the numerous uncertain model parameters within various physical, biological, and chemical processes that influence methane production, oxidation, and transport. Sensitivity Analysis (SA) can help identify critical parameters for methane emission and achieve reduced biases and uncertainties in future projections. This study performs SA for 19 selected parameters responsible for critical biogeochemical processes in the methane module of the Energy Exascale Earth System Model (E3SM) land model (ELM). The impact of these parameters on various CH4 fluxes is examined at 14 FLUXNET- CH4 sites with diverse vegetation types. Given the extensive number of model simulations needed for global variance-based SA, we employ a machine learning (ML) algorithm to emulate the complex behavior of ELM methane biogeochemistry. ML enables the computational time to be shortened significantly from 6 CPU hours to 0.72 milliseconds, achieving reduced computational costs. We found that parameters linked to CH4 production and diffusion generally present the highest sensitivities despite apparent seasonal variation. Comparing simulated emissions from perturbed parameter sets against FLUXNET-CH4 observations revealed that better performances can be achieved at each site compared to the default parameter values. This presents a scope for further improving simulated emissions using parameter calibration with advanced optimization techniques like Bayesian optimization.
We provide high-speed implementations for simulating reservoirs described by $N$-coupled spin-torque oscillators. Here $N$ also corresponds to the number of reservoir nodes. We benchmark a variety of implementations based on CPU and GPU. Our new methods are at least 2.6 times quicker than the baseline for $N$ in range $1$ to $10^4$. More specifically, over all implementations the best factor is 78.9 for $N=1$ which decreases to 2.6 for $N=10^3$ and finally increases to 23.8 for $N=10^4$. GPU outperforms CPU significantly at $N=2500$. Our results show that GPU implementations should be tested for reservoir simulations. The implementations considered here can be used for any reservoir with evolution that can be approximated using an explicit method.
Transformer-based trackers have achieved strong accuracy on the standard benchmarks. However, their efficiency remains an obstacle to practical deployment on both GPU and CPU platforms. In this paper, to overcome this issue, we propose a fully transformer tracking framework, coined as \emph{MixFormerV2}, without any dense convolutional operation and complex score prediction module. Our key design is to introduce four special prediction tokens and concatenate them with the tokens from target template and search areas. Then, we apply the unified transformer backbone on these mixed token sequence. These prediction tokens are able to capture the complex correlation between target template and search area via mixed attentions. Based on them, we can easily predict the tracking box and estimate its confidence score through simple MLP heads. To further improve the efficiency of MixFormerV2, we present a new distillation-based model reduction paradigm, including dense-to-sparse distillation and deep-to-shallow distillation. The former one aims to transfer knowledge from the dense-head based MixViT to our fully transformer tracker, while the latter one is used to prune some layers of the backbone. We instantiate two types of MixForemrV2, where the MixFormerV2-B achieves an AUC of 70.6\% on LaSOT and an AUC of 57.4\% on TNL2k with a high GPU speed of 165 FPS, and the MixFormerV2-S surpasses FEAR-L by 2.7\% AUC on LaSOT with a real-time CPU speed.
The bi-encoder architecture provides a framework for understanding machine-learned retrieval models based on dense and sparse vector representations. Although these representations capture parametric realizations of the same underlying conceptual framework, their respective implementations of top-$k$ similarity search require the coordination of different software components (e.g., inverted indexes, HNSW indexes, and toolkits for neural inference), often knitted together in complex architectures. In this work, we ask the following question: What's the simplest design, in terms of requiring the fewest changes to existing infrastructure, that can support end-to-end retrieval with modern dense and sparse representations? The answer appears to be that Lucene is sufficient, as we demonstrate in Anserini, a toolkit for reproducible information retrieval research. That is, effective retrieval with modern single-vector neural models can be efficiently performed directly in Java on the CPU. We examine the implications of this design for information retrieval researchers pushing the state of the art as well as for software engineers building production search systems.
In this paper, we investigate how to push the performance limits of serving Deep Neural Network (DNN) models on CPU-based servers. Specifically, we observe that while intra-operator parallelism across multiple threads is an effective way to reduce inference latency, it provides diminishing returns. Our primary insight is that instead of running a single instance of a model with all available threads on a server, running multiple instances each with smaller batch sizes and fewer threads for intra-op parallelism can provide lower inference latency. However, the right configuration is hard to determine manually since it is workload- (DNN model and batch size used by the serving system) and deployment-dependent (number of CPU cores on server). We present Packrat, a new serving system for online inference that given a model and batch size ($B$) algorithmically picks the optimal number of instances ($i$), the number of threads each should be allocated ($t$), and the batch sizes each should operate on ($b$) that minimizes latency. Packrat is built as an extension to TorchServe and supports online reconfigurations to avoid serving downtime. Averaged across a range of batch sizes, Packrat improves inference latency by 1.43$\times$ to 1.83$\times$ on a range of commonly used DNNs.
Since the proposal of transformers, these models have been limited to bounded input lengths, because of their need to attend to every token in the input. In this work, we propose Unlimiformer: a general approach that wraps any existing pretrained encoder-decoder transformer, and offloads the cross-attention computation to a single k-nearest-neighbor (kNN) index, while the returned kNN distances are the attention dot-product scores. This kNN index can be kept on either the GPU or CPU memory and queried in sub-linear time; this way, we can index practically unlimited input sequences, while every attention head in every decoder layer retrieves its top-k keys, instead of attending to every key. We evaluate Unlimiformer on several long-document and book-summarization benchmarks, showing that it can process even 500k token-long inputs from the BookSum dataset, without any input truncation at test time. We demonstrate that Unlimiformer improves pretrained models such as BART and Longformer by extending them to unlimited inputs without additional learned weights and without modifying their code. We make our code and models publicly available at https://github.com/abertsch72/unlimiformer .
Graph Neural Networks (GNNs) play a crucial role in various fields. However, most existing deep graph learning frameworks assume pre-stored static graphs and do not support training on graph streams. In contrast, many real-world graphs are dynamic and contain time domain information. We introduce GNNFlow, a distributed framework that enables efficient continuous temporal graph representation learning on dynamic graphs on multi-GPU machines. GNNFlow introduces an adaptive time-indexed block-based data structure that effectively balances memory usage with graph update and sampling operation efficiency. It features a hybrid GPU-CPU graph data placement for rapid GPU-based temporal neighborhood sampling and kernel optimizations for enhanced sampling processes. A dynamic GPU cache for node and edge features is developed to maximize cache hit rates through reuse and restoration strategies. GNNFlow supports distributed training across multiple machines with static scheduling to ensure load balance. We implement GNNFlow based on DGL and PyTorch. Our experimental results show that GNNFlow provides up to 21.1x faster continuous learning than existing systems.
Support Vector Machine (SVM) algorithm requires a high computational cost (both in memory and time) to solve a complex quadratic programming (QP) optimization problem during the training process. Consequently, SVM necessitates high computing hardware capabilities. The central processing unit (CPU) clock frequency cannot be increased due to physical limitations in the miniaturization process. However, the potential of parallel multi-architecture, available in both multi-core CPUs and highly scalable GPUs, emerges as a promising solution to enhance algorithm performance. Therefore, there is an opportunity to reduce the high computational time required by SVM for solving the QP optimization problem. This paper presents a comparative study that implements the SVM algorithm on different parallel architecture frameworks. The experimental results show that SVM MPI-CUDA implementation achieves a speedup over SVM TensorFlow implementation on different datasets. Moreover, SVM TensorFlow implementation provides a cross-platform solution that can be migrated to alternative hardware components, which will reduces the development time.
The lattice Boltzmann method (LBM) has emerged as a prominent technique for solving fluid dynamics problems due to its algorithmic potential for computational scalability. We introduce XLB library, a Python-based differentiable LBM library based on the JAX platform. The architecture of XLB is predicated upon ensuring accessibility, extensibility, and computational performance, enabling scaling effectively across CPU, TPU, multi-GPU, and distributed multi-GPU or TPU systems. The library can be readily augmented with novel boundary conditions, collision models, or multi-physics simulation capabilities. XLB's differentiability and data structure is compatible with the extensive JAX-based machine learning ecosystem, enabling it to address physics-based machine learning, optimization, and inverse problems. XLB has been successfully scaled to handle simulations with billions of cells, achieving giga-scale lattice updates per second. XLB is released under the permissive Apache-2.0 license and is available on GitHub at https://github.com/Autodesk/XLB.
Kidney stone disease results in millions of annual visits to emergency departments in the United States. Computed tomography (CT) scans serve as the standard imaging modality for efficient detection of kidney stones. Various approaches utilizing convolutional neural networks (CNNs) have been proposed to implement automatic diagnosis of kidney stones. However, there is a growing interest in employing fast and efficient CNNs on edge devices in clinical practice. In this paper, we propose a lightweight fusion framework for kidney detection and kidney stone diagnosis on coronal CT images. In our design, we aim to minimize the computational costs of training and inference while implementing an automated approach. The experimental results indicate that our framework can achieve competitive outcomes using only 8\% of the original training data. These results include an F1 score of 96\% and a False Negative (FN) error rate of 4\%. Additionally, the average detection time per CT image on a CPU is 0.62 seconds. Reproducibility: Framework implementation and models available on GitHub.
Recent quantization techniques have enabled heterogeneous precisions at very fine granularity, e.g., each parameter/activation can take on a different precision, resulting in compact neural networks without sacrificing accuracy. However, there is a lack of efficient architectural support for such networks, which require additional hardware to decode the precision settings for individual variables, align the variables, and provide fine-grained mixed-precision compute capabilities. The complexity of these operations introduces high overheads. Thus, the improvements in inference latency/energy of these networks are not commensurate with the compression ratio, and may be inferior to larger quantized networks with uniform precisions. We present an end-to-end co-design approach encompassing computer architecture, training algorithm, and inference optimization to efficiently execute networks with fine-grained heterogeneous precisions. The key to our approach is a novel training algorithm designed to accommodate hardware constraints and inference operation requirements, outputting networks with input-channel-wise heterogeneous precisions and at most three precision levels. Combined with inference optimization techniques, existing architectures with low-cost enhancements can support such networks efficiently, yielding optimized tradeoffs between accuracy, compression ratio and inference latency/energy. We demonstrate the efficacy of our approach across CPU and GPU architectures. For various representative neural networks, our approach achieves >10x improvements in both compression ratio and inference latency, with negligible degradation in accuracy compared to full-precision networks.
Graph Neural Networks (GNNs) have demonstrated outstanding performance in various applications. Existing frameworks utilize CPU-GPU heterogeneous environments to train GNN models and integrate mini-batch and sampling techniques to overcome the GPU memory limitation. In CPU-GPU heterogeneous environments, we can divide sample-based GNN training into three steps: sample, gather, and train. Existing GNN systems use different task orchestrating methods to employ each step on CPU or GPU. After extensive experiments and analysis, we find that existing task orchestrating methods fail to fully utilize the heterogeneous resources, limited by inefficient CPU processing or GPU resource contention. In this paper, we propose NeutronOrch, a system for sample-based GNN training that incorporates a layer-based task orchestrating method and ensures balanced utilization of the CPU and GPU. NeutronOrch decouples the training process by layer and pushes down the training task of the bottom layer to the CPU. This significantly reduces the computational load and memory footprint of GPU training. To avoid inefficient CPU processing, NeutronOrch only offloads the training of frequently accessed vertices to the CPU and lets GPU reuse their embeddings with bounded staleness. Furthermore, NeutronOrch provides a fine-grained pipeline design for the layer-based task orchestrating method, fully overlapping different tasks on heterogeneous resources while strictly guaranteeing bounded staleness. The experimental results show that compared with the state-of-the-art GNN systems, NeutronOrch can achieve up to 11.51x performance speedup.
Transformer-based large language models (LLMs) have demonstrated impressive capabilities in a variety of natural language processing (NLP) tasks. Nonetheless, it is challenging to deploy and fine-tune LLMs on mobile edge devices with limited computing, memory, and energy budgets. In this paper, we propose Confidant, a multi-backend collaborative training framework for customizing state-of-the-art LLMs on commodity mobile devices like smartphones. Confidant partitions an LLM into several sub-models so that each fits into a mobile device's memory. A pipeline parallel training mechanism is further developed to ensure fast and efficient distributed training. In addition, we propose a novel backend scheduler to allocate different attention heads to heterogeneous compute hardware, including mobile CPU and GPUs, to maximize the compute resource utilization on each edge device. Our preliminary experimental results show that Confidant achieves at most 45.3% memory reduction and 8.03x inference speedup in practical settings.
This paper presents a novel approach to neuromorphic audio processing by integrating the strengths of Spiking Neural Networks (SNNs), Transformers, and high-performance computing (HPC) into the HPCNeuroNet architecture. Utilizing the Intel N-DNS dataset, we demonstrate the system's capability to process diverse human vocal recordings across multiple languages and noise backgrounds. The core of our approach lies in the fusion of the temporal dynamics of SNNs with the attention mechanisms of Transformers, enabling the model to capture intricate audio patterns and relationships. Our architecture, HPCNeuroNet, employs the Short-Time Fourier Transform (STFT) for time-frequency representation, Transformer embeddings for dense vector generation, and SNN encoding/decoding mechanisms for spike train conversions. The system's performance is further enhanced by leveraging the computational capabilities of NVIDIA's GeForce RTX 3060 GPU and Intel's Core i9 12900H CPU. Additionally, we introduce a hardware implementation on the Xilinx VU37P HBM FPGA platform, optimizing for energy efficiency and real-time processing. The proposed accelerator achieves a throughput of 71.11 Giga-Operations Per Second (GOP/s) with a 3.55 W on-chip power consumption at 100 MHz. The comparison results with off-the-shelf devices and recent state-of-the-art implementations illustrate that the proposed accelerator has obvious advantages in terms of energy efficiency and design flexibility. Through design-space exploration, we provide insights into optimizing core capacities for audio tasks. Our findings underscore the transformative potential of integrating SNNs, Transformers, and HPC for neuromorphic audio processing, setting a new benchmark for future research and applications.
Unsupervised environment design (UED) is a form of automatic curriculum learning for training robust decision-making agents to zero-shot transfer into unseen environments. Such autocurricula have received much interest from the RL community. However, UED experiments, based on CPU rollouts and GPU model updates, have often required several weeks of training. This compute requirement is a major obstacle to rapid innovation for the field. This work introduces the minimax library for UED training on accelerated hardware. Using JAX to implement fully-tensorized environments and autocurriculum algorithms, minimax allows the entire training loop to be compiled for hardware acceleration. To provide a petri dish for rapid experimentation, minimax includes a tensorized grid-world based on MiniGrid, in addition to reusable abstractions for conducting autocurricula in procedurally-generated environments. With these components, minimax provides strong UED baselines, including new parallelized variants, which achieve over 120$\times$ speedups in wall time compared to previous implementations when training with equal batch sizes. The minimax library is available under the Apache 2.0 license at https://github.com/facebookresearch/minimax.
Linear system solving is a key tool for computational power system studies, e.g., optimal power flow, transmission switching, or unit commitment. CPU-based linear system solver speeds, however, have saturated in recent years. Emerging research shows that GPU-based linear system solvers are beginning to achieve notable speedup over CPU-based alternatives in some applications. Due to the architecture of GPU memory access, numerical pivoting represents the new bottleneck which prevents GPU-based solvers from running even faster. Accordingly, this paper proposes a matrix perturbation-based method to induce static pivoting. Using this approach, a series of perturbed, well-conditioned, pivot-free linear systems are solved in parallel on GPUs. Matrix expansion routines are then used to linearly combine the results, and the true solution is recovered to an arbitrarily high degree of theoretical accuracy. We showcase the validity of our approach on distributed-slack AC power flow solve iterations associated with the PGLib 300-bus test case.
Deep Neural Networks (DNNs) have been widely used for illumination estimation, which is time-consuming and requires sensor-specific data collection. Our proposed method uses a dual-mapping strategy and only requires a simple white point from a test sensor under a D65 condition. This allows us to derive a mapping matrix, enabling the reconstructions of image data and illuminants. In the second mapping phase, we transform the re-constructed image data into sparse features, which are then optimized with a lightweight multi-layer perceptron (MLP) model using the re-constructed illuminants as ground truths. This approach effectively reduces sensor discrepancies and delivers performance on par with leading cross-sensor methods. It only requires a small amount of memory (~0.003 MB), and takes ~1 hour training on an RTX3070Ti GPU. More importantly, the method can be implemented very fast, with ~0.3 ms and ~1 ms on a GPU or CPU respectively, and is not sensitive to the input image resolution. Therefore, it offers a practical solution to the great challenges of data recollection that is faced by the industry.
Language models only really need to use an exponential fraction of their neurons for individual inferences. As proof, we present UltraFastBERT, a BERT variant that uses 0.3% of its neurons during inference while performing on par with similar BERT models. UltraFastBERT selectively engages just 12 out of 4095 neurons for each layer inference. This is achieved by replacing feedforward networks with fast feedforward networks (FFFs). While no truly efficient implementation currently exists to unlock the full acceleration potential of conditional neural execution, we provide high-level CPU code achieving 78x speedup over the optimized baseline feedforward implementation, and a PyTorch implementation delivering 40x speedup over the equivalent batched feedforward inference. We publish our training code, benchmarking setup, and model weights.
Background and objectives: Dynamic handwriting analysis, due to its non-invasive and readily accessible nature, has recently emerged as a vital adjunctive method for the early diagnosis of Parkinson's disease. In this study, we design a compact and efficient network architecture to analyse the distinctive handwriting patterns of patients' dynamic handwriting signals, thereby providing an objective identification for the Parkinson's disease diagnosis. Methods: The proposed network is based on a hybrid deep learning approach that fully leverages the advantages of both long short-term memory (LSTM) and convolutional neural networks (CNNs). Specifically, the LSTM block is adopted to extract the time-varying features, while the CNN-based block is implemented using one-dimensional convolution for low computational cost. Moreover, the hybrid model architecture is continuously refined under ablation studies for superior performance. Finally, we evaluate the proposed method with its generalization under a five-fold cross-validation, which validates its efficiency and robustness. Results: The proposed network demonstrates its versatility by achieving impressive classification accuracies on both our new DraWritePD dataset ($96.2\%$) and the well-established PaHaW dataset ($90.7\%$). Moreover, the network architecture also stands out for its excellent lightweight design, occupying a mere $0.084$M of parameters, with a total of only $0.59$M floating-point operations. It also exhibits near real-time CPU inference performance, with inference times ranging from $0.106$ to $0.220$s. Conclusions: We present a series of experiments with extensive analysis, which systematically demonstrate the effectiveness and efficiency of the proposed hybrid neural network in extracting distinctive handwriting patterns for precise diagnosis of Parkinson's disease.