Representing and rendering dynamic scenes has been an important but challenging task. Especially, to accurately model complex motions, high efficiency is usually hard to guarantee. To achieve real-time dynamic scene rendering while also enjoying high training and storage efficiency, we propose 4D Gaussian Splatting (4D-GS) as a holistic representation for dynamic scenes rather than applying 3D-GS for each individual frame. In 4D-GS, a novel explicit representation containing both 3D Gaussians and 4D neural voxels is proposed. A decomposed neural voxel encoding algorithm inspired by HexPlane is proposed to efficiently build Gaussian features from 4D neural voxels and then a lightweight MLP is applied to predict Gaussian deformations at novel timestamps. Our 4D-GS method achieves real-time rendering under high resolutions, 82 FPS at an 800$\times$800 resolution on an RTX 3090 GPU while maintaining comparable or better quality than previous state-of-the-art methods. More demos and code are available at https://guanjunwu.github.io/4dgs/.
Large Language Models (LLMs) excel in NLP, but their demands hinder their widespread deployment. While Quantization-Aware Training (QAT) offers a solution, its extensive training costs make Post-Training Quantization (PTQ) a more practical approach for LLMs. In existing studies, activation outliers in particular channels are identified as the bottleneck to PTQ accuracy. They propose to transform the magnitudes from activations to weights, which however offers limited alleviation or suffers from unstable gradients, resulting in a severe performance drop at low-bitwidth. In this paper, we propose QLLM, an accurate and efficient low-bitwidth PTQ method designed for LLMs. QLLM introduces an adaptive channel reassembly technique that reallocates the magnitude of outliers to other channels, thereby mitigating their impact on the quantization range. This is achieved by channel disassembly and channel assembly, which first breaks down the outlier channels into several sub-channels to ensure a more balanced distribution of activation magnitudes. Then similar channels are merged to maintain the original channel number for efficiency. Additionally, an adaptive strategy is designed to autonomously determine the optimal number of sub-channels for channel disassembly. To further compensate for the performance loss caused by quantization, we propose an efficient tuning method that only learns a small number of low-rank weights while freezing the pre-trained quantized model. After training, these low-rank parameters can be fused into the frozen weights without affecting inference. Extensive experiments on LLaMA-1 and LLaMA-2 show that QLLM can obtain accurate quantized models efficiently. For example, QLLM quantizes the 4-bit LLaMA-2-70B within 10 hours on a single A100-80G GPU, outperforming the previous state-of-the-art method by 7.89% on the average accuracy across five zero-shot tasks.
During construction, continuous monitoring of underground tunnels can mitigate potential hazards and facilitate an in-depth understanding of the ground-tunnel interaction behavior. Traditional vision-based monitoring can directly capture an extensive range of motion but cannot separate the tunnel's vibration and deformation mode. Phase-based motion magnification is one of the techniques to magnify the motion in target frequency bands and identify system dynamics. Optical flow is a popular method of calculating the motion of image intensities in computer vision and has a much lower computational cost than Digital Image Correlation. This study combines PMM and OF to quantify the underground tunnel scene's magnified deformation mode pixel displacements. As motion magnification artifacts may lead to inaccurate quantification, the 2D Wiener filter is used to smooth the high-frequency content. With GPU acceleration, a dense OF algorithm computing each pixel's displacement is adopted to derive the whole scene motion. A validation experiment is conducted between the amplification motion and the actual motion of prisms preinstalled in the tunnel.
We consider the problem of accurate sparse fine-tuning of large language models (LLMs), that is, fine-tuning pretrained LLMs on specialized tasks, while inducing sparsity in their weights. On the accuracy side, we observe that standard loss-based fine-tuning may fail to recover accuracy, especially at high sparsities. To address this, we perform a detailed study of distillation-type losses, determining an L2-based distillation approach we term SquareHead which enables accurate recovery even at higher sparsities, across all model types. On the practical efficiency side, we show that sparse LLMs can be executed with speedups by taking advantage of sparsity, for both CPU and GPU runtimes. While the standard approach is to leverage sparsity for computational reduction, we observe that in the case of memory-bound LLMs sparsity can also be leveraged for reducing memory bandwidth. We exhibit end-to-end results showing speedups due to sparsity, while recovering accuracy, on T5 (language translation), Whisper (speech translation), and open GPT-type (MPT for text generation). For MPT text generation, we show for the first time that sparse fine-tuning can reach 75% sparsity without accuracy drops, provide notable end-to-end speedups for both CPU and GPU inference, and highlight that sparsity is also compatible with quantization approaches. Models and software for reproducing our results are provided in Section 6.
On-device ML introduces new security challenges: DNN models become white-box accessible to device users. Based on white-box information, adversaries can conduct effective model stealing (MS) and membership inference attack (MIA). Using Trusted Execution Environments (TEEs) to shield on-device DNN models aims to downgrade (easy) white-box attacks to (harder) black-box attacks. However, one major shortcoming is the sharply increased latency (up to 50X). To accelerate TEE-shield DNN computation with GPUs, researchers proposed several model partition techniques. These solutions, referred to as TEE-Shielded DNN Partition (TSDP), partition a DNN model into two parts, offloading the privacy-insensitive part to the GPU while shielding the privacy-sensitive part within the TEE. This paper benchmarks existing TSDP solutions using both MS and MIA across a variety of DNN models, datasets, and metrics. We show important findings that existing TSDP solutions are vulnerable to privacy-stealing attacks and are not as safe as commonly believed. We also unveil the inherent difficulty in deciding optimal DNN partition configurations (i.e., the highest security with minimal utility cost) for present TSDP solutions. The experiments show that such ``sweet spot'' configurations vary across datasets and models. Based on lessons harvested from the experiments, we present TEESlice, a novel TSDP method that defends against MS and MIA during DNN inference. TEESlice follows a partition-before-training strategy, which allows for accurate separation between privacy-related weights from public weights. TEESlice delivers the same security protection as shielding the entire DNN model inside TEE (the ``upper-bound'' security guarantees) with over 10X less overhead (in both experimental and real-world environments) than prior TSDP solutions and no accuracy loss.
Large Language Models (LLMs) have showcased remarkable impacts across a wide spectrum of natural language processing tasks. Fine-tuning these pre-trained models on downstream datasets provides further significant performance gains, but this process has been challenging due to its extraordinary resource requirements. To this end, existing efforts focus on parameter-efficient fine-tuning, which, unfortunately, fail to capitalize on the powerful potential of full-parameter fine-tuning. In this work, we propose QFT, a novel Quantized Full-parameter Tuning framework for LLMs that enables memory-efficient fine-tuning without harming performance. Our framework incorporates two novel ideas: (i) we adopt the efficient Lion optimizer, which only keeps track of the momentum and has consistent update magnitudes for each parameter, an inherent advantage for robust quantization; and (ii) we quantize all model states and store them as integer values, and present a gradient flow and parameter update scheme for the quantized weights. As a result, QFT reduces the model state memory to 21% of the standard solution while achieving comparable performance, e.g., tuning a LLaMA-7B model requires only <30GB of memory, satisfied by a single A6000 GPU.
Pretraining auto-regressive large language models~(LLMs) with retrieval demonstrates better perplexity and factual accuracy by leveraging external databases. However, the size of existing pretrained retrieval-augmented LLM is still limited (e.g., Retro has 7.5B parameters), which limits the effectiveness of instruction tuning and zero-shot generalization. In this work, we introduce Retro 48B, the largest LLM pretrained with retrieval. Specifically, we continue to pretrain a 43B GPT model on additional 100 billion tokens using the Retro augmentation method by retrieving from 1.2 trillion tokens. Notably, the obtained foundation model, Retro 48B, largely outperforms the counterpart GPT 43B trained on 1.2T tokens in terms of perplexity with only 2.58% additional GPU hours, demonstrating the significant scaling potential of the method. After instruction tuning on Retro, InstructRetro demonstrates significant improvement over the instruction tuned GPT on a wide range of zero-shot tasks. Specifically, the average improvement of InstructRetro is 7% over its GPT counterpart across 8 short-form QA and reading comprehension tasks, 10% over GPT across 4 challenging long-form QA tasks, and 16% over GPT across 3 summarization tasks. Surprisingly, we find that one can ablate the encoder from InstructRetro architecture and directly use its decoder backbone, while achieving comparable results. Our results highlight the promising direction to obtain a better GPT decoder through continued pretraining with retrieval before instruction tuning. Our code and checkpoints are publicly available at: https://huggingface.co/nvidia/retro-48b-instruct-4k.
Semantic communication aims to transmit meaningful and effective information, rather than focusing on individual symbols or bits. This results in benefits like reduced latency, bandwidth usage, and higher throughput compared with traditional communication. However, semantic communication poses significant challenges due to the need for universal metrics to benchmark the joint effects of semantic information loss and practical energy consumption. This research presents a novel multi-objective loss function named "Energy-Optimized Semantic Loss" (EOSL), addressing the challenge of balancing semantic information loss and energy consumption. Through comprehensive experiments on transformer models, including CPU and GPU energy usage, it is demonstrated that EOSL-based encoder model selection can save up to 90% of energy while achieving a 44% improvement in semantic similarity performance during inference in this experiment. This work paves the way for energy-efficient neural network selection and the development of greener semantic communication architectures.
In this paper, we explore how transfer learning, coupled with Intel Xeon, specifically 4th Gen Intel Xeon scalable processor, defies the conventional belief that training is primarily GPU-dependent. We present a case study where we achieved near state-of-the-art accuracy for image classification on a publicly available Image Classification TensorFlow dataset using Intel Advanced Matrix Extensions(AMX) and distributed training with Horovod.
AI models are increasing in size and recent advancement in the community has shown that unlike HPC applications where double precision datatype are required, lower-precision datatypes such as fp8 or int4 are sufficient to bring the same model quality both for training and inference. Following these trends, GPU vendors such as NVIDIA and AMD have added hardware support for fp16, fp8 and int8 GeMM operations with an exceptional performance via Tensor Cores. However, this paper proposes a new algorithm called msGeMM which shows that AI models with low-precision datatypes can run with ~2.5x fewer multiplication and add instructions. Efficient implementation of this algorithm requires special CUDA cores with the ability to add elements from a small look-up table at the rate of Tensor Cores.
AI accelerator processing capabilities and memory constraints largely dictate the scale in which machine learning workloads (e.g., training and inference) can be executed within a desirable time frame. Training a state of the art, transformer-based model today requires use of GPU-accelerated high performance computers with high-speed interconnects. As datasets and models continue to increase in size, computational requirements and memory demands for AI also continue to grow. These challenges have inspired the development of distributed algorithm and circuit-based optimization techniques that enable the ability to progressively scale models in multi-node environments, efficiently minimize neural network cost functions for faster convergence, and store more parameters into a set number of available resources. In our research project, we focus on parallel and distributed machine learning algorithm development, specifically for optimizing the data processing and pre-training of a set of 5 encoder-decoder LLMs, ranging from 580 million parameters to 13 billion parameters. We performed a fine-grained study to quantify the relationships between three ML parallelism methods, specifically exploring Microsoft DeepSpeed Zero Redundancy Optimizer (ZeRO) stages.
Recent advances in robot-assisted surgery have resulted in progressively more precise, efficient, and minimally invasive procedures, sparking a new era of robotic surgical intervention. This enables doctors, in collaborative interaction with robots, to perform traditional or minimally invasive surgeries with improved outcomes through smaller incisions. Recent efforts are working toward making robotic surgery more autonomous which has the potential to reduce variability of surgical outcomes and reduce complication rates. Deep reinforcement learning methodologies offer scalable solutions for surgical automation, but their effectiveness relies on extensive data acquisition due to the absence of prior knowledge in successfully accomplishing tasks. Due to the intensive nature of simulated data collection, previous works have focused on making existing algorithms more efficient. In this work, we focus on making the simulator more efficient, making training data much more accessible than previously possible. We introduce Surgical Gym, an open-source high performance platform for surgical robot learning where both the physics simulation and reinforcement learning occur directly on the GPU. We demonstrate between 100-5000x faster training times compared with previous surgical learning platforms. The code is available at: https://github.com/SamuelSchmidgall/SurgicalGym.
Estimating tissue parameter maps with high accuracy and precision from highly undersampled measurements presents one of the major challenges in MR fingerprinting (MRF). Many existing works project the recovered voxel fingerprints onto the Bloch manifold to improve reconstruction performance. However, little research focuses on exploiting the latent manifold structure priors among fingerprints. To fill this gap, we propose a novel MRF reconstruction framework based on manifold structured data priors. Since it is difficult to directly estimate the fingerprint manifold structure, we model the tissue parameters as points on a low-dimensional parameter manifold. We reveal that the fingerprint manifold shares the same intrinsic topology as the parameter manifold, although being embedded in different Euclidean spaces. To exploit the non-linear and non-local redundancies in MRF data, we divide the MRF data into spatial patches, and the similarity measurement among data patches can be accurately obtained using the Euclidean distance between the corresponding patches in the parameter manifold. The measured similarity is then used to construct the graph Laplacian operator, which represents the fingerprint manifold structure. Thus, the fingerprint manifold structure is introduced in the reconstruction framework by using the low-dimensional parameter manifold. Additionally, we incorporate the locally low-rank prior in the reconstruction framework to further utilize the local correlations within each patch for improved reconstruction performance. We also adopt a GPU-accelerated NUFFT library to accelerate reconstruction in non-Cartesian sampling scenarios. Experimental results demonstrate that our method can achieve significantly improved reconstruction performance with reduced computational time over the state-of-the-art methods.
What happens if we encounter a suitable font for our design work but do not know its name? Visual Font Recognition (VFR) systems are used to identify the font typeface in an image. These systems can assist graphic designers in identifying fonts used in images. A VFR system also aids in improving the speed and accuracy of Optical Character Recognition (OCR) systems. In this paper, we introduce the first publicly available datasets in the field of Persian font recognition and employ Convolutional Neural Networks (CNN) to address this problem. The results show that the proposed pipeline obtained 78.0% top-1 accuracy on our new datasets, 89.1% on the IDPL-PFOD dataset, and 94.5% on the KAFD dataset. Furthermore, the average time spent in the entire pipeline for one sample of our proposed datasets is 0.54 and 0.017 seconds for CPU and GPU, respectively. We conclude that CNN methods can be used to recognize Persian fonts without the need for additional pre-processing steps such as feature extraction, binarization, normalization, etc.
This paper introduces a distributed, GPU-centric experience replay system, GEAR, designed to perform scalable reinforcement learning (RL) with large sequence models (such as transformers). With such models, existing systems such as Reverb face considerable bottlenecks in memory, computation, and communication. GEAR, however, optimizes memory efficiency by enabling the memory resources on GPU servers (including host memory and device memory) to manage trajectory data. Furthermore, it facilitates decentralized GPU devices to expedite various trajectory selection strategies, circumventing computational bottlenecks. GEAR is equipped with GPU kernels capable of collecting trajectories using zero-copy access to host memory, along with remote-directed-memory access over InfiniBand, improving communication efficiency. Cluster experiments have shown that GEAR can achieve performance levels up to 6x greater than Reverb when training state-of-the-art large RL models. GEAR is open-sourced at https://github.com/bigrl-team/gear.
Amidst the swift advancements in photography and sensor technologies, high-definition cameras have become commonplace in the deployment of Unmanned Aerial Vehicles (UAVs) for diverse operational purposes. Within the domain of UAV imagery analysis, the segmentation of ultra-high resolution images emerges as a substantial and intricate challenge, especially when grappling with the constraints imposed by GPU memory-restricted computational devices. This paper delves into the intricate problem of achieving efficient and effective segmentation of ultra-high resolution UAV imagery, while operating under stringent GPU memory limitation. The strategy of existing approaches is to downscale the images to achieve computationally efficient segmentation. However, this strategy tends to overlook smaller, thinner, and curvilinear regions. To address this problem, we propose a GPU memory-efficient and effective framework for local inference without accessing the context beyond local patches. In particular, we introduce a novel spatial-guided high-resolution query module, which predicts pixel-wise segmentation results with high quality only by querying nearest latent embeddings with the guidance of high-resolution information. Additionally, we present an efficient memory-based interaction scheme to correct potential semantic bias of the underlying high-resolution information by associating cross-image contextual semantics. For evaluation of our approach, we perform comprehensive experiments over public benchmarks and achieve superior performance under both conditions of small and large GPU memory usage limitations. We will release the model and codes in the future.
Objective: The objective of this work is an attempt towards non-contact freehand 3D ultrasound imaging with minimal complexity added to the existing point of care ultrasound (POCUS) systems. Methods: This study proposes a novel approach of using a mechanical track for non-contact ultrasound (US) scanning. The approach thus restricts the probe motion to a linear plane, to simplify the acquisition and 3D reconstruction process. A pipeline for US 3D volume reconstruction employing an US research platform and a GPU-based edge device is developed. Results: The efficacy of the proposed approach is demonstrated through ex-vivo and in-vivo experiments. Conclusion: The proposed approach with the adjustable field of view capability, non-contact design, and low cost of deployment without significantly altering the existing setup would open doors for up gradation of traditional systems to a wide range of 3D US imaging applications. Significance: Ultrasound (US) imaging is a popular clinical imaging modality for the point-of-care bedside imaging, particularly of the wrist/knee in the pediatric population due to its non-invasive and radiation free nature. However, the limited views of tissue structures obtained with 2D US in such scenarios make the diagnosis challenging. To overcome this, 3D US imaging which uses 2D US images and their orientation/position to reconstruct 3D volumes was developed. The accurate position estimation of the US probe at low cost has always stood as a challenging task in 3D reconstruction. Additionally, US imaging involves contact, which causes difficulty to pediatric subjects while monitoring live fractures or open wounds. Towards overcoming these challenges, a novel framework is attempted in this work.
Latent Diffusion models (LDMs) have achieved remarkable results in synthesizing high-resolution images. However, the iterative sampling process is computationally intensive and leads to slow generation. Inspired by Consistency Models (song et al.), we propose Latent Consistency Models (LCMs), enabling swift inference with minimal steps on any pre-trained LDMs, including Stable Diffusion (rombach et al). Viewing the guided reverse diffusion process as solving an augmented probability flow ODE (PF-ODE), LCMs are designed to directly predict the solution of such ODE in latent space, mitigating the need for numerous iterations and allowing rapid, high-fidelity sampling. Efficiently distilled from pre-trained classifier-free guided diffusion models, a high-quality 768 x 768 2~4-step LCM takes only 32 A100 GPU hours for training. Furthermore, we introduce Latent Consistency Fine-tuning (LCF), a novel method that is tailored for fine-tuning LCMs on customized image datasets. Evaluation on the LAION-5B-Aesthetics dataset demonstrates that LCMs achieve state-of-the-art text-to-image generation performance with few-step inference. Project Page: https://latent-consistency-models.github.io/
Neural Networks design is a complex and often daunting task, particularly for resource-constrained scenarios typical of mobile-sized models. Neural Architecture Search is a promising approach to automate this process, but existing competitive methods require large training time and computational resources to generate accurate models. To overcome these limits, this paper contributes with: i) a novel training-free metric, named Entropic Score, to estimate model expressivity through the aggregated element-wise entropy of its activations; ii) a cyclic search algorithm to separately yet synergistically search model size and topology. Entropic Score shows remarkable ability in searching for the topology of the network, and a proper combination with LogSynflow, to search for model size, yields superior capability to completely design high-performance Hybrid Transformers for edge applications in less than 1 GPU hour, resulting in the fastest and most accurate NAS method for ImageNet classification.
Convolutional Neural Networks (CNNs) have produced state-of-the-art results for image classification tasks. However, they are limited in their ability to handle rotational and viewpoint variations due to information loss in max-pooling layers. Capsule Networks (CapsNets) employ a computationally-expensive iterative process referred to as dynamic routing to address these issues. CapsNets, however, often fall short on complex datasets and require more computational resources than CNNs. To overcome these challenges, we introduce the Parallel Dynamic Routing CapsNet (PDR-CapsNet), a deeper and more energy-efficient alternative to CapsNet that offers superior performance, less energy consumption, and lower overfitting rates. By leveraging a parallelization strategy, PDR-CapsNet mitigates the computational complexity of CapsNet and increases throughput, efficiently using hardware resources. As a result, we achieve 83.55\% accuracy while requiring 87.26\% fewer parameters, 32.27\% and 47.40\% fewer MACs, and Flops, achieving 3x faster inference and 7.29J less energy consumption on a 2080Ti GPU with 11GB VRAM compared to CapsNet and for the CIFAR-10 dataset.