Accurate and efficient catheter segmentation in 3D ultrasound (US) is essential for cardiac intervention. Currently, the state-of-the-art segmentation algorithms are based on convolutional neural networks (CNNs), which achieved remarkable performances in a standard Cartesian volumetric data. Nevertheless, these approaches suffer the challenges of low efficiency and GPU unfriendly image size. Therefore, such difficulties and expensive hardware requirements become a bottleneck to build accurate and efficient segmentation models for real clinical application. In this paper, we propose a novel Frustum ultrasound based catheter segmentation method. Specifically, Frustum ultrasound is a polar coordinate based image, which includes same information of standard Cartesian image but has much smaller size, which overcomes the bottleneck of efficiency than conventional Cartesian images. Nevertheless, the irregular and deformed Frustum images lead to more efforts for accurate voxel-level annotation. To address this limitation, a weakly supervised learning framework is proposed, which only needs 3D bounding box annotations overlaying the region-of-interest to training the CNNs. Although the bounding box annotation includes noise and inaccurate annotation to mislead to model, it is addressed by the proposed pseudo label generated scheme. The labels of training voxels are generated by incorporating class activation maps with line filtering, which is iteratively updated during the training. Our experimental results show the proposed method achieved the state-of-the-art performance with an efficiency of 0.25 second per volume. More crucially, the Frustum image segmentation provides a much faster and cheaper solution for segmentation in 3D US image, which meet the demands of clinical applications.
Flame Spray Pyrolysis (FSP) is a manufacturing technique to mass produce engineered nanoparticles for applications in catalysis, energy materials, composites, and more. FSP instruments are highly dependent on a number of adjustable parameters, including fuel injection rate, fuel-oxygen mixtures, and temperature, which can greatly affect the quality, quantity, and properties of the yielded nanoparticles. Optimizing FSP synthesis requires monitoring, analyzing, characterizing, and modifying experimental conditions.Here, we propose a hybrid CPU-GPU Difference of Gaussians (DoG)method for characterizing the volume distribution of unburnt solution, so as to enable near-real-time optimization and steering of FSP experiments. Comparisons against standard implementations show our method to be an order of magnitude more efficient. This surrogate signal can be deployed as a component of an online end-to-end pipeline that maximizes the synthesis yield.
Differentiable Architecture Search (DARTS) provides a baseline for searching effective network architectures based gradient, but it is accompanied by huge computational overhead in searching and training network architecture. Recently, many novel works have improved DARTS. Particularly, Partially-Connected DARTS(PC-DARTS) proposed the partial channel sampling technique which achieved good results. In this work, we found that the backbone provided by DARTS is prone to overfitting. To mitigate this problem, we propose an approach named Group-DARTS with Attention (G-DARTS-A), using multiple groups of channels for searching. Inspired by the partially sampling strategy of PC-DARTS, we use groups channels to sample the super-network to perform a more efficient search while maintaining the relative integrity of the network information. In order to relieve the competition between channel groups and keep channel balance, we follow the attention mechanism in Squeeze-and-Excitation Network. Each group of channels shares defined weights thence they can provide different suggestion for searching. The searched architecture is more powerful and better adapted to different deployments. Specifically, by only using the attention module on DARTS we achieved an error rate of 2.82%/16.36% on CIFAR10/100 with 0.3GPU-days for search process on CIFAR10. Apply our G-DARTS-A to DARTS/PC-DARTS, an error rate of 2.57%/2.61% on CIFAR10 with 0.5/0.4 GPU-days is achieved.
Hamiltonian Monte Carlo (HMC) is a Markov chain Monte Carlo (MCMC) approach that exhibits favourable exploration properties in high-dimensional models such as neural networks. Unfortunately, HMC has limited use in large-data regimes and little work has explored suitable approaches that aim to preserve the entire Hamiltonian. In our work, we introduce a new symmetric integration scheme for split HMC that does not rely on stochastic gradients. We show that our new formulation is more efficient than previous approaches and is easy to implement with a single GPU. As a result, we are able to perform full HMC over common deep learning architectures using entire data sets. In addition, when we compare with stochastic gradient MCMC, we show that our method achieves better performance in both accuracy and uncertainty quantification. Our approach demonstrates HMC as a feasible option when considering inference schemes for large-scale machine learning problems.
The tool voxlogica merges the state-of-the-art library of computational imaging algorithms ITK with the combination of declarative specification and optimised execution provided by spatial logic model checking. The analysis of an existing benchmark for segmentation of brain tumours via a simple logical specification reached state-of-the-art accuracy. We present a new, GPU-based version of voxlogica and discuss its implementation, scalability, and applications.
Automatic and consistent meningioma segmentation in T1-weighted MRI volumes and corresponding volumetric assessment is of use for diagnosis, treatment planning, and tumor growth evaluation. In this paper, we optimized the segmentation and processing speed performances using a large number of both surgically treated meningiomas and untreated meningiomas followed at the outpatient clinic. We studied two different 3D neural network architectures: (i) a simple encoder-decoder similar to a 3D U-Net, and (ii) a lightweight multi-scale architecture (PLS-Net). In addition, we studied the impact of different training schemes. For the validation studies, we used 698 T1-weighted MR volumes from St. Olav University Hospital, Trondheim, Norway. The models were evaluated in terms of detection accuracy, segmentation accuracy and training/inference speed. While both architectures reached a similar Dice score of 70% on average, the PLS-Net was more accurate with an F1-score of up to 88%. The highest accuracy was achieved for the largest meningiomas. Speed-wise, the PLS-Net architecture tended to converge in about 50 hours while 130 hours were necessary for U-Net. Inference with PLS-Net takes less than a second on GPU and about 15 seconds on CPU. Overall, with the use of mixed precision training, it was possible to train competitive segmentation models in a relatively short amount of time using the lightweight PLS-Net architecture. In the future, the focus should be brought toward the segmentation of small meningiomas (less than 2ml) to improve clinical relevance for automatic and early diagnosis as well as speed of growth estimates.
Learning system dynamics directly from observations is a promising direction in machine learning due to its potential to significantly enhance our ability to understand physical systems. However, the dynamics of many real-world systems are challenging to learn due to the presence of nonlinear potentials and a number of interactions that scales quadratically with the number of particles $N$, as in the case of the N-body problem. In this work, we introduce an approach that transforms a fully-connected interaction graph into a hierarchical one which reduces the number of edges to $O(N)$. This results in linear time and space complexity while the pre-computation of the hierarchical graph requires $O(N\log (N))$ time and $O(N)$ space. Using our approach, we are able to train models on much larger particle counts, even on a single GPU. We evaluate how the phase space position accuracy and energy conservation depend on the number of simulated particles. Our approach retains high accuracy and efficiency even on large-scale gravitational N-body simulations which are impossible to run on a single machine if a fully-connected graph is used. Similar results are also observed when simulating Coulomb interactions. Furthermore, we make several important observations regarding the performance of this new hierarchical model, including: i) its accuracy tends to improve with the number of particles in the simulation and ii) its generalisation to unseen particle counts is also much better than for models that use all $O(N^2)$ interactions.
Neural architecture search (NAS) aims to produce the optimal sparse solution from a high-dimensional space spanned by all candidate connections. Current gradient-based NAS methods commonly ignore the constraint of sparsity in the search phase, but project the optimized solution onto a sparse one by post-processing. As a result, the dense super-net for search is inefficient to train and has a gap with the projected architecture for evaluation. In this paper, we formulate neural architecture search as a sparse coding problem. We perform the differentiable search on a compressed lower-dimensional space that has the same validation loss as the original sparse solution space, and recover an architecture by solving the sparse coding problem. The differentiable search and architecture recovery are optimized in an alternate manner. By doing so, our network for search at each update satisfies the sparsity constraint and is efficient to train. In order to also eliminate the depth and width gap between the network in search and the target-net in evaluation, we further propose a method to search and evaluate in one stage under the target-net settings. When training finishes, architecture variables are absorbed into network weights. Thus we get the searched architecture and optimized parameters in a single run. In experiments, our two-stage method on CIFAR-10 requires only 0.05 GPU-day for search. Our one-stage method produces state-of-the-art performances on both CIFAR-10 and ImageNet at the cost of only evaluation time.
The transformer is the most critical algorithm innovation of the Nature Language Processing (NLP) field in recent years. Unlike the Recurrent Neural Network (RNN) models, Transformers can process on dimensions of sequence lengths in parallel, therefore leading to better accuracy on long sequences. However, efficient deployments of them for online services in data centers equipped with GPUs are not easy. First, more computation introduced by transformer structures makes it more challenging to meet the latency and throughput constraints of serving. Second, NLP tasks take in sentences of variable length. The variability of input dimensions brings a severe problem to efficient memory management and serving optimization. This paper designed a transformer serving system called TurboTransformers, which consists of a computing runtime and a serving framework to solve the above challenges. Three innovative features make it stand out from other similar works. An efficient parallel algorithm is proposed for GPU-based batch reduction operations, like Softmax and LayerNorm, major hot spots besides BLAS routines. A memory allocation algorithm, which better balances the memory footprint and allocation/free efficiency, is designed for variable-length input situations. A serving framework equipped with a new batch scheduler using dynamic programming achieves the optimal throughput on variable-length requests. The system can achieve the state-of-the-art transformer model serving performance on GPU platforms and can be seamlessly integrated into your PyTorch code with a few lines of code.
Machine Learning (ML) adoption in the enterprise requires simpler and more efficient software infrastructure---the bespoke solutions typical in large web companies are simply untenable. Model scoring, the process of obtaining predictions from a trained model over new data, is a primary contributor to infrastructure complexity and cost as models are trained once but used many times. In this paper we propose HUMMINGBIRD, a novel approach to model scoring, which compiles featurization operators and traditional ML models (e.g., decision trees) into a small set of tensor operations. This approach inherently reduces infrastructure complexity and directly leverages existing investments in Neural Network compilers and runtimes to generate efficient computations for both CPU and hardware accelerators. Our performance results are intriguing: despite replacing imperative computations (e.g., tree traversals) with tensor computation abstractions, HUMMINGBIRD is competitive and often outperforms hand-crafted kernels on micro-benchmarks on both CPU and GPU, while enabling seamless end-to-end acceleration of ML pipelines. We have released HUMMINGBIRD as open source.
Several recent work on speech synthesis have employed generative adversarial networks (GANs) to produce raw waveforms. Although such methods improve the sampling efficiency and memory usage, their sample quality has not yet reached that of autoregressive and flow-based generative models. In this work, we propose HiFi-GAN, which achieves both efficient and high-fidelity speech synthesis. As speech audio consists of sinusoidal signals with various periods, we demonstrate that modeling periodic patterns of an audio is crucial for enhancing sample quality. A subjective human evaluation (mean opinion score, MOS) of a single speaker dataset indicates that our proposed method demonstrates similarity to human quality while generating 22.05 kHz high-fidelity audio 167.9 times faster than real-time on a single V100 GPU. We further show the generality of HiFi-GAN to the mel-spectrogram inversion of unseen speakers and end-to-end speech synthesis. Finally, a small footprint version of HiFi-GAN generates samples 13.4 times faster than real-time on CPU with comparable quality to an autoregressive counterpart.
Speaker recognition is an active research area that contains notable usage in biometric security and authentication system. Currently, there exist many well-performing models in the speaker recognition domain. However, most of the advanced models implement deep learning that requires GPU support for real-time speech recognition, and it is not suitable for low-end devices. In this paper, we propose a lightweight text-independent speaker recognition model based on random forest classifier. It also introduces new features that are used for both speaker verification and identification tasks. The proposed model uses human speech based timbral properties as features that are classified using random forest. Timbre refers to the very basic properties of sound that allow listeners to discriminate among them. The prototype uses seven most actively searched timbre properties, boominess, brightness, depth, hardness, roughness, sharpness, and warmth as features of our speaker recognition model. The experiment is carried out on speaker verification and speaker identification tasks and shows the achievements and drawbacks of the proposed model. In the speaker identification phase, it achieves a maximum accuracy of 78%. On the contrary, in the speaker verification phase, the model maintains an accuracy of 80% having an equal error rate (ERR) of 0.24.
Face recognition has been an active and vital topic among computer vision community for a long time. Previous researches mainly focus on loss functions used for facial feature extraction network, among which the improvements of softmax-based loss functions greatly promote the performance of face recognition. However, the contradiction between the drastically increasing number of face identities and the shortage of GPU memories is gradually becoming irreconcilable. In this paper, we thoroughly analyze the optimization goal of softmax-based loss functions and the difficulty of training massive identities. We find that the importance of negative classes in softmax function in face representation learning is not as high as we previously thought. The experiment demonstrates no loss of accuracy when training with only 10\% randomly sampled classes for the softmax-based loss functions, compared with training with full classes using state-of-the-art models on mainstream benchmarks. We also implement a very efficient distributed sampling algorithm, taking into account model accuracy and training efficiency, which uses only eight NVIDIA RTX2080Ti to complete classification tasks with tens of millions of identities. The code of this paper has been made available https://github.com/deepinsight/insightface/tree/master/recognition/partial_fc.
Accurate characterization of glioma is crucial for clinical decision making. A delineation of the tumor is also desirable in the initial decision stages but is a time-consuming task. Leveraging the latest GPU capabilities, we developed a single multi-task convolutional neural network that uses the full 3D, structural, pre-operative MRI scans to can predict the IDH mutation status, the 1p/19q co-deletion status, and the grade of a tumor, while simultaneously segmenting the tumor. We trained our method using the largest, most diverse patient cohort to date containing 1508 glioma patients from 16 institutes. We tested our method on an independent dataset of 240 patients from 13 different institutes, and achieved an IDH-AUC of 0.90, 1p/19q-AUC of 0.85, grade-AUC of 0.81, and a mean whole tumor DICE score of 0.84. Thus, our method non-invasively predicts multiple, clinically relevant parameters and generalizes well to the broader clinical population.
We propose a unified appearance model accounting for traditional shallow (i.e. 3D SIFT keypoints) and deep (i.e. CNN output layers) image feature representations, encoding respectively specific, localized neuroanatomical patterns and rich global information into a single indexing and classification framework. A novel Bayesian model combines shallow and deep features based on an assumption of conditional independence and validated by experiments indexing specific family members and general group categories in 3D MRI neuroimage data of 1010 subjects from the Human Connectome Project, including twins and non-twin siblings. A novel domain adaptation strategy is presented, transforming deep CNN vectors elements into binary class-informative descriptors. A GPU-based implementation of all processing is provided. State-of-the-art performance is achieved in large-scale neuroimage indexing, both in terms of computational complexity, accuracy in identifying family members and sex classification.
Tensor decompositions, such as CANDECOMP/PARAFAC (CP), are widely used in a variety of applications, such as chemometrics, signal processing, and machine learning. A broadly used method for computing such decompositions relies on the Alternating Least Squares (ALS) algorithm. When the number of components is small, regardless of its implementation, ALS exhibits low arithmetic intensity, which severely hinders its performance and makes GPU offloading ineffective. We observe that, in practice, experts often have to compute multiple decompositions of the same tensor, each with a small number of components (typically fewer than 20), to ultimately find the best ones to use for the application at hand. In this paper, we illustrate how multiple decompositions of the same tensor can be fused together at the algorithmic level to increase the arithmetic intensity. Therefore, it becomes possible to make efficient use of GPUs for further speedups; at the same time the technique is compatible with many enhancements typically used in ALS, such as line search, extrapolation, and non-negativity constraints. We introduce the Concurrent ALS algorithm and library, which offers an interface to Matlab, and a mechanism to effectively deal with the issue that decompositions complete at different times. Experimental results on artificial and real datasets demonstrate a shorter time to completion due to increased arithmetic intensity.
In the last few years, several deep learning models, especially Generative Adversarial Networks have received a lot of attention for the task of Single Image Super-Resolution (SISR). These methods focus on building an end-to-end framework, which produce a high resolution(SR) image from a given low resolution(LR) image in a single step to achieve state-of-the-art performance. This paper focuses on improving an existing deep-learning based method to perform Super-Resolution Microscopy in real-time using a standard GPU. For this, we first propose a tiling strategy, which takes advantage of parallelism provided by a GPU to speed up the network training process. Further, we suggest simple changes to the architecture of the generator and the discriminator of SRGAN. Subsequently, We compare the quality and the running time for the outputs produced by our model, opening its applications in different areas like low-end benchtop and even mobile microscopy. Finally, we explore the possibility of the trained network to produce High-Resolution HR outputs for different domains.
Modern deep learning systems like PyTorch and Tensorflow are able to train enormous models with billions (or trillions) of parameters on a distributed infrastructure. These systems require that the internal nodes have the same memory capacity and compute performance. Unfortunately, most organizations, especially universities, have a piecemeal approach to purchasing computer systems resulting in a heterogeneous infrastructure, which cannot be used to compute large models. The present work describes HetSeq, a software package adapted from the popular PyTorch package that provides the capability to train large neural network models on heterogeneous infrastructure. Experiments with transformer translation and BERT language model shows that HetSeq scales over heterogeneous systems. HetSeq can be easily extended to other models like image classification. Package with supported document is publicly available at https://github.com/yifding/hetseq.
TensorBNN is a new package based on TensorFlow that implements Bayesian inference for modern neural network models. The posterior density of neural network model parameters is represented as a point cloud sampled using Hamiltonian Monte Carlo. The TensorBNN package leverages TensorFlow's architecture and training features as well as its ability to use modern graphics processing units (GPU) in both the training and prediction stages.
Generative adversarial networks (GANs) are currently rarely applied on 3D medical images of large size, due to their immense computational demand. The present work proposes a multi-scale patch-based GAN approach for establishing unpaired domain translation by generating 3D medical image volumes of high resolution in a memory-efficient way. The key idea to enable memory-efficient image generation is to first generate a low-resolution version of the image followed by the generation of patches of constant sizes but successively growing resolutions. To avoid patch artifacts and incorporate global information, the patch generation is conditioned on patches from previous resolution scales. Those multi-scale GANs are trained to generate realistically looking images from image sketches in order to perform an unpaired domain translation. This allows to preserve the topology of the test data and generate the appearance of the training domain data. The evaluation of the domain translation scenarios is performed on brain MRIs of size 155x240x240 and thorax CTs of size up to 512x512x512. Compared to common patch-based approaches, the multi-resolution scheme enables better image quality and prevents patch artifacts. Also, it ensures constant GPU memory demand independent from the image size, allowing for the generation of arbitrarily large images.