This work describes a process for efficiently fine-tuning the GraphCast data-driven forecast model to simulate another analysis system, here the Global Deterministic Prediction System (GDPS) of Environment and Climate Change Canada (ECCC). Using two years of training data (July 2019 -- December 2021) and 37 GPU-days of computation to tune the 37-level, quarter-degree version of GraphCast, the resulting model significantly outperforms both the unmodified GraphCast and operational forecast, showing significant forecast skill in the troposphere over lead times from 1 to 10 days. This fine-tuning is accomplished through abbreviating DeepMind's original training curriculum for GraphCast, relying on a shorter single-step forecast stage to accomplish the bulk of the adaptation work and consolidating the autoregressive stages into separate 12hr, 1d, 2d, and 3d stages with larger learning rates. Additionally, training over 3d forecasts is split into two sub-steps to conserve host memory while maintaining a strong correlation with training over the full period.
In medical contexts, the imbalanced data distribution in long-tailed datasets, due to scarce labels for rare diseases, greatly impairs the diagnostic accuracy of deep learning models. Recent multimodal text-image supervised foundation models offer new solutions to data scarcity through effective representation learning. However, their limited medical-specific pretraining hinders their performance in medical image classification relative to natural images. To address this issue, we propose a novel Text-guided Foundation model Adaptation for Long-Tailed medical image classification (TFA-LT). We adopt a two-stage training strategy, integrating representations from the foundation model using just two linear adapters and a single ensembler for balanced outcomes. Experimental results on two long-tailed medical image datasets validate the simplicity, lightweight and efficiency of our approach: requiring only 6.1% GPU memory usage of the current best-performing algorithm, our method achieves an accuracy improvement of up to 27.1%, highlighting the substantial potential of foundation model adaptation in this area.
Diffusion models have emerged as a popular method for 3D generation. However, it is still challenging for diffusion models to efficiently generate diverse and high-quality 3D shapes. In this paper, we introduce OctFusion, which can generate 3D shapes with arbitrary resolutions in 2.5 seconds on a single Nvidia 4090 GPU, and the extracted meshes are guaranteed to be continuous and manifold. The key components of OctFusion are the octree-based latent representation and the accompanying diffusion models. The representation combines the benefits of both implicit neural representations and explicit spatial octrees and is learned with an octree-based variational autoencoder. The proposed diffusion model is a unified multi-scale U-Net that enables weights and computation sharing across different octree levels and avoids the complexity of widely used cascaded diffusion schemes. We verify the effectiveness of OctFusion on the ShapeNet and Objaverse datasets and achieve state-of-the-art performances on shape generation tasks. We demonstrate that OctFusion is extendable and flexible by generating high-quality color fields for textured mesh generation and high-quality 3D shapes conditioned on text prompts, sketches, or category labels. Our code and pre-trained models are available at https://github.com/octree-nn/octfusion.
Linear RNN architectures, like Mamba, can be competitive with Transformer models in language modeling while having advantageous deployment characteristics. Given the focus on training large-scale Transformer models, we consider the challenge of converting these pretrained models for deployment. We demonstrate that it is feasible to distill large Transformers into linear RNNs by reusing the linear projection weights from attention layers with academic GPU resources. The resulting hybrid model, which incorporates a quarter of the attention layers, achieves performance comparable to the original Transformer in chat benchmarks and outperforms open-source hybrid Mamba models trained from scratch with trillions of tokens in both chat benchmarks and general benchmarks. Moreover, we introduce a hardware-aware speculative decoding algorithm that accelerates the inference speed of Mamba and hybrid models. Overall we show how, with limited computation resources, we can remove many of the original attention layers and generate from the resulting model more efficiently. Our top-performing model, distilled from Llama3-8B-Instruct, achieves a 29.61 length-controlled win rate on AlpacaEval 2 against GPT-4 and 7.35 on MT-Bench, surpassing the best 8B scale instruction-tuned linear RNN model. We also find that the distilled model has natural length extrapolation, showing almost perfect accuracy in the needle-in-a-haystack test at 20x the distillation length. Code and pre-trained checkpoints are open-sourced at https://github.com/jxiw/MambaInLlama and https://github.com/itsdaniele/speculative_mamba.
Counterfactual regret minimization is a family of algorithms of no-regret learning dynamics capable of solving large-scale imperfect information games. We propose implementing this algorithm as a series of dense and sparse matrix and vector operations, thereby making it highly parallelizable for a graphical processing unit, at a cost of higher memory usage. Our experiments show that our implementation performs up to about 401.2 times faster than OpenSpiel's Python implementation and, on an expanded set of games, up to about 203.6 times faster than OpenSpiel's C++ implementation and the speedup becomes more pronounced as the size of the game being solved grows.
Learning from label proportions (LLP) is a kind of weakly supervised learning that trains an instance-level classifier from label proportions of bags, which consist of sets of instances without using instance labels. A challenge in LLP arises when the number of instances in a bag (bag size) is numerous, making the traditional LLP methods difficult due to GPU memory limitations. This study aims to develop an LLP method capable of learning from bags with large sizes. In our method, smaller bags (mini-bags) are generated by sampling instances from large-sized bags (original bags), and these mini-bags are used in place of the original bags. However, the proportion of a mini-bag is unknown and differs from that of the original bag, leading to overfitting. To address this issue, we propose a perturbation method for the proportion labels of sampled mini-bags to mitigate overfitting to noisy label proportions. This perturbation is added based on the multivariate hypergeometric distribution, which is statistically modeled. Additionally, loss weighting is implemented to reduce the negative impact of proportions sampled from the tail of the distribution. Experimental results demonstrate that the proportion label perturbation and loss weighting achieve classification accuracy comparable to that obtained without sampling. Our codes are available at https://github.com/stainlessnight/LLP-LargeBags.
Open Set Object Detection has seen rapid development recently, but it continues to pose significant challenges. Language-based methods, grappling with the substantial modal disparity between textual and visual modalities, require extensive computational resources to bridge this gap. Although integrating visual prompts into these frameworks shows promise for enhancing performance, it always comes with constraints related to textual semantics. In contrast, viusal-only methods suffer from the low-quality fusion of multiple visual prompts. In response, we introduce a strong DETR-based model, Visual Intersection Network for Open Set Object Detection (VINO), which constructs a multi-image visual bank to preserve the semantic intersections of each category across all time steps. Our innovative multi-image visual updating mechanism learns to identify the semantic intersections from various visual prompts, enabling the flexible incorporation of new information and continuous optimization of feature representations. Our approach guarantees a more precise alignment between target category semantics and region semantics, while significantly reducing pre-training time and resource demands compared to language-based methods. Furthermore, the integration of a segmentation head illustrates the broad applicability of visual intersection in various visual tasks. VINO, which requires only 7 RTX4090 GPU days to complete one epoch on the Objects365v1 dataset, achieves competitive performance on par with vision-language models on benchmarks such as LVIS and ODinW35.
Pillar-based 3D object detection has gained traction in self-driving technology due to its speed and accuracy facilitated by the artificial densification of pillars for GPU-friendly processing. However, dense pillar processing fundamentally wastes computation since it ignores the inherent sparsity of pillars derived from scattered point cloud data. Motivated by recent embedded accelerators with native sparsity support, sparse pillar convolution methods like submanifold convolution (SubM-Conv) aimed to reduce these redundant computations by applying convolution only on active pillars but suffered considerable accuracy loss. Our research identifies that this accuracy loss is due to the restricted fine-grained spatial information flow (fSIF) of SubM-Conv in sparse pillar networks. To overcome this restriction, we propose a selectively dilated (SD-Conv) convolution that evaluates the importance of encoded pillars and selectively dilates the convolution output, enhancing the receptive field for critical pillars and improving object detection accuracy. To facilitate actual acceleration with this novel convolution approach, we designed SPADE+ as a cost-efficient augmentation to existing embedded sparse convolution accelerators. This design supports the SD-Conv without significant demands in area and SRAM size, realizing superior trade-off between the speedup and model accuracy. This strategic enhancement allows our method to achieve extreme pillar sparsity, leading to up to 18.1x computational savings and 16.2x speedup on the embedded accelerators, without compromising object detection accuracy.
Fourier Ptychographic Microscopy (FPM) is a computational imaging technique that enables high-resolution imaging over a large field of view. However, its application in the biomedical field has been limited due to the long image reconstruction time and poor noise robustness. In this paper, we propose a fast and robust FPM reconstruction method based on physical neural networks with batch update stochastic gradient descent (SGD) optimization strategy, capable of achieving attractive results with low single-to-noise ratio and correcting multiple system parameters simultaneously. Our method leverages a random batch optimization approach, breaks away from the fixed sequential iterative order and gives greater attention to high-frequency information. The proposed method has better convergence performance even for low signal-to-noise ratio data sets, such as low exposure time dark-field images. As a result, it can greatly increase the image recording and result reconstruction speed without any additional hardware modifications. By utilizing advanced deep learning optimizers and perform parallel computational scheme, our method enhances GPU computational efficiency, significantly reducing reconstruction costs. Experimental results demonstrate that our method achieves near real-time digital refocusing of a 1024 x 1024 pixels region of interest on consumer-grade GPUs. This approach significantly improves temporal resolution (by reducing the exposure time of dark-field images), noise resistance, and reconstruction speed, and therefore can efficiently promote the practical application of FPM in clinical diagnostics, digital pathology, and biomedical research, etc. In addition, we believe our algorithm scheme can help researchers quickly validate and implement FPM-related ideas. We invite requests for the full code via email.
Multi-GPU nodes are increasingly common in the rapidly evolving landscape of exascale supercomputers. On these systems, GPUs on the same node are connected through dedicated networks, with bandwidths up to a few terabits per second. However, gauging performance expectations and maximizing system efficiency is challenging due to different technologies, design options, and software layers. This paper comprehensively characterizes three supercomputers - Alps, Leonardo, and LUMI - each with a unique architecture and design. We focus on performance evaluation of intra-node and inter-node interconnects on up to 4096 GPUs, using a mix of intra-node and inter-node benchmarks. By analyzing its limitations and opportunities, we aim to offer practical guidance to researchers, system architects, and software developers dealing with multi-GPU supercomputing. Our results show that there is untapped bandwidth, and there are still many opportunities for optimization, ranging from network to software optimization.
Dynamic scene reconstruction using Gaussians has recently attracted increased interest. Mainstream approaches typically employ a global deformation field to warp a 3D scene in canonical space. However, the inherent low-frequency nature of implicit neural fields often leads to ineffective representations of complex motions. Moreover, their structural rigidity can hinder adaptation to scenes with varying resolutions and durations. To address these challenges, we introduce a novel approach for streaming 4D real-world reconstruction utilizing discrete 3D control points. This method physically models local rays and establishes a motion-decoupling coordinate system. By effectively merging traditional graphics with learnable pipelines, it provides a robust and efficient local 6-degrees-of-freedom (6-DoF) motion representation. Additionally, we have developed a generalized framework that integrates our control points with Gaussians. Starting from an initial 3D reconstruction, our workflow decomposes the streaming 4D reconstruction into four independent submodules: 3D segmentation, 3D control point generation, object-wise motion manipulation, and residual compensation. Experimental results demonstrate that our method outperforms existing state-of-the-art 4D Gaussian splatting techniques on both the Neu3DV and CMU-Panoptic datasets. Notably, the optimization of our 3D control points is achievable in 100 iterations and within just 2 seconds per frame on a single NVIDIA 4070 GPU.
Implementing AI algorithms on event-based embedded devices enables real-time processing of data, minimizes latency, and enhances power efficiency in edge computing. This research explores the deployment of a spiking recurrent neural network (SRNN) with liquid time constant neurons for gesture recognition. We focus on the energy efficiency and computational efficacy of NVIDIA Jetson Nano embedded GPU platforms. The embedded GPU showcases a 14-fold increase in power efficiency relative to a conventional GPU, making a compelling argument for its use in energy-constrained applications. The study's empirical findings also highlight that batch processing significantly boosts frame rates across various batch sizes while maintaining accuracy levels well above the baseline. These insights validate the SRNN with liquid time constant neurons as a robust model for interpreting temporal-spatial data in gesture recognition, striking a critical balance between processing speed and power frugality.
Graph Neural Networks (GNNs) are becoming increasingly popular for graph-based learning tasks such as point cloud processing due to their state-of-the-art (SOTA) performance. Nevertheless, the research community has primarily focused on improving model expressiveness, lacking consideration of how to design efficient GNN models for edge scenarios with real-time requirements and limited resources. Examining existing GNN models reveals varied execution across platforms and frequent Out-Of-Memory (OOM) problems, highlighting the need for hardware-aware GNN design. To address this challenge, this work proposes a novel hardware-aware graph neural architecture search framework tailored for resource constraint edge devices, namely HGNAS. To achieve hardware awareness, HGNAS integrates an efficient GNN hardware performance predictor that evaluates the latency and peak memory usage of GNNs in milliseconds. Meanwhile, we study GNN memory usage during inference and offer a peak memory estimation method, enhancing the robustness of architecture evaluations when combined with predictor outcomes. Furthermore, HGNAS constructs a fine-grained design space to enable the exploration of extreme performance architectures by decoupling the GNN paradigm. In addition, the multi-stage hierarchical search strategy is leveraged to facilitate the navigation of huge candidates, which can reduce the single search time to a few GPU hours. To the best of our knowledge, HGNAS is the first automated GNN design framework for edge devices, and also the first work to achieve hardware awareness of GNNs across different platforms. Extensive experiments across various applications and edge devices have proven the superiority of HGNAS. It can achieve up to a 10.6x speedup and an 82.5% peak memory reduction with negligible accuracy loss compared to DGCNN on ModelNet40.
This paper presents a novel and fast approach for ground plane segmentation in a LiDAR point cloud, specifically optimized for processing speed and hardware efficiency on FPGA hardware platforms. Our approach leverages a channel-based segmentation method with an advanced angular data repair technique and a cross-eight-way flood-fill algorithm. This innovative approach significantly reduces the number of iterations while ensuring the high accuracy of the segmented ground plane, which makes the stream-based hardware implementation possible. To validate the proposed approach, we conducted extensive experiments on the SemanticKITTI dataset. We introduced a bird's-eye view (BEV) evaluation metric tailored for the area representation of LiDAR segmentation tasks. Our method demonstrated superior performance in terms of BEV areas when compared to the existing approaches. Moreover, we presented an optimized hardware architecture targeted on a Zynq-7000 FPGA, compatible with LiDARs of various channel densities, i.e., 32, 64, and 128 channels. Our FPGA implementation operating at 160 MHz significantly outperforms the traditional computing platforms, which is 12 to 25 times faster than the CPU-based solutions and up to 6 times faster than the GPU-based solution, in addition to the benefit of low power consumption.
Due to high accuracy, BERT-like models have been widely adopted by discriminative text mining and web searching. However, large BERT-like models suffer from inefficient online inference, as they face the following two problems on GPUs. First, they rely on the large model depth to achieve high accuracy, which linearly increases the sequential computation on GPUs. Second, stochastic and dynamic online workloads cause extra costs. In this paper, we present Academus for low-latency online inference of BERT-like models. At the core of Academus is the novel student parallelism, which adopts boosting ensemble and stacking distillation to distill the original deep model into an equivalent group of parallel and shallow student models. This enables Academus to achieve the lower model depth (e.g., two layers) than baselines and consequently the lowest inference latency without affecting the accuracy.For occasional workload bursts, it can temporarily decrease the number of students with minimal accuracy loss to improve throughput. Additionally, it employs specialized system designs for student parallelism to better handle stochastic online workloads. We conduct comprehensive experiments to verify the effectiveness. The results show that Academus outperforms the baselines by 4.1X~1.6X in latency without compromising accuracy, and achieves up to 22.27X higher throughput for workload bursts.
Recent advances in mapping techniques have enabled the creation of highly accurate dense 3D maps during robotic missions, such as point clouds, meshes, or NeRF-based representations. These developments present new opportunities for reusing these maps for localization. However, there remains a lack of a unified approach that can operate seamlessly across different map representations. This paper presents and evaluates a global visual localization system capable of localizing a single camera image across various 3D map representations built using both visual and lidar sensing. Our system generates a database by synthesizing novel views of the scene, creating RGB and depth image pairs. Leveraging the precise 3D geometric map, our method automatically defines rendering poses, reducing the number of database images while preserving retrieval performance. To bridge the domain gap between real query camera images and synthetic database images, our approach utilizes learning-based descriptors and feature detectors. We evaluate the system's performance through extensive real-world experiments conducted in both indoor and outdoor settings, assessing the effectiveness of each map representation and demonstrating its advantages over traditional structure-from-motion (SfM) localization approaches. The results show that all three map representations can achieve consistent localization success rates of 55% and higher across various environments. NeRF synthesized images show superior performance, localizing query images at an average success rate of 72%. Furthermore, we demonstrate an advantage over SfM-based approaches that our synthesized database enables localization in the reverse travel direction which is unseen during the mapping process. Our system, operating in real-time on a mobile laptop equipped with a GPU, achieves a processing rate of 1Hz.
Procedural Content Generation via Reinforcement Learning (PCGRL) has been introduced as a means by which controllable designer agents can be trained based only on a set of computable metrics acting as a proxy for the level's quality and key characteristics. While PCGRL offers a unique set of affordances for game designers, it is constrained by the compute-intensive process of training RL agents, and has so far been limited to generating relatively small levels. To address this issue of scale, we implement several PCGRL environments in Jax so that all aspects of learning and simulation happen in parallel on the GPU, resulting in faster environment simulation; removing the CPU-GPU transfer of information bottleneck during RL training; and ultimately resulting in significantly improved training speed. We replicate several key results from prior works in this new framework, letting models train for much longer than previously studied, and evaluating their behavior after 1 billion timesteps. Aiming for greater control for human designers, we introduce randomized level sizes and frozen "pinpoints" of pivotal game tiles as further ways of countering overfitting. To test the generalization ability of learned generators, we evaluate models on large, out-of-distribution map sizes, and find that partial observation sizes learn more robust design strategies.
Deep learning has become a popular tool across many scientific fields, including the study of differential equations, particularly partial differential equations. This work introduces the basic principles of deep learning and the Deep Galerkin method, which uses deep neural networks to solve differential equations. This primer aims to provide technical and practical insights into the Deep Galerkin method and its implementation. We demonstrate how to solve the one-dimensional heat equation step-by-step. We also show how to apply the Deep Galerkin method to solve systems of ordinary differential equations and integral equations, such as the Fredholm of the second kind. Additionally, we provide code snippets within the text and the complete source code on Github. The examples are designed so that one can run them on a simple computer without needing a GPU.
In this work, we present a GPU-accelerated library for the underlying components of Kolmogorov-Arnold Networks (KANs), along with an algorithm to eliminate bounded grids in KANs. The GPU-accelerated library reduces the computational complexity of Basis Spline (B-spline) evaluation by a factor of $\mathcal{O}$(grid size) compared to existing codes, enabling batch computation for large-scale learning. To overcome the limitations of traditional KANs, we introduce Unbounded KANs (UKANs), which eliminate the need for a bounded grid and a fixed number of B-spline coefficients. To do so, we replace the KAN parameters (B-spline coefficients) with a coefficient generator (CG) model. The inputs to the CG model are designed based on the idea of an infinite symmetric grid extending from negative infinity to positive infinity. The positional encoding of grid group, a sequential collection of B-spline grid indexes, is fed into the CG model, and coefficients are consumed by the efficient implementation (matrix representations) of B-spline functions to generate outputs. We perform several experiments on regression, classification, and generative tasks, which are promising. In particular, UKAN does not require data normalization or a bounded domain for evaluation. Additionally, our benchmarking results indicate the superior memory and computational efficiency of our library compared to existing codes.
This paper compared physics-informed neural network (PINN), conventional neural network (NN) and traditional numerical discretization methods on solving differential equations (DEs) through literature investigation and experimental validation. We focused on the soft-constrained PINN approach and formalized its mathematical framework and computational flow for solving Ordinary DEs and Partial DEs (ODEs/PDEs). The working mechanism and its accuracy and efficiency were experimentally verified by solving typical linear and non-linear oscillator ODEs. We demonstrate that the DeepXDE-based implementation of PINN is not only light code and efficient in training, but also flexible across CPU/GPU platforms. PINN greatly reduces the need for labeled data: when the nonlinearity of the ODE is weak, a very small amount of supervised training data plus a few unsupervised collocation points are sufficient to predict the solution; in the minimalist case, only one or two training points (with initial values) are needed for first- or second-order ODEs, respectively. We also find that, with the aid of collocation points and the use of physical information, PINN has the ability to extrapolate data outside the time domain of the training set, and especially is robust to noisy data, thus with enhanced generalization capabilities. Training is accelerated when the gains obtained along with the reduction in the amount of data outweigh the delay caused by the increase in the loss function terms. The soft-constrained PINN can easily impose a physical law (e.g., conservation of energy) constraint by adding a regularization term to the total loss function, thus improving the solution performance to ODEs that obey this physical law. Furthermore, PINN can also be used for stiff ODEs, PDEs, and other types of DEs, and is becoming a favorable catalyst for the era of Digital Twins.