Automatic neural architecture design has shown its potential in discovering powerful neural network architectures. Existing methods, no matter based on reinforcement learning or evolutionary algorithms (EA), conduct architecture search in a discrete space, which is highly inefficient. In this paper, we propose a simple and efficient method to automatic neural architecture design based on continuous optimization. We call this new approach neural architecture optimization (NAO). There are three key components in our proposed approach: (1) An encoder embeds/maps neural network architectures into a continuous space. (2) A predictor takes the continuous representation of a network as input and predicts its accuracy. (3) A decoder maps a continuous representation of a network back to its architecture. The performance predictor and the encoder enable us to perform gradient based optimization in the continuous space to find the embedding of a new architecture with potentially better accuracy. Such a better embedding is then decoded to a network by the decoder. Experiments show that the architecture discovered by our method is very competitive for image classification task on CIFAR-10 and language modeling task on PTB, outperforming or on par with the best results of previous architecture search methods with a significantly reduction of computational resources. Specifically we obtain 1.93% test set error rate for CIFAR-10 image classification task and 56.0 test set perplexity of PTB language modeling task. Furthermore, combined with the recent proposed weight sharing mechanism, we discover powerful architecture on CIFAR-10 (with error rate 2.93%) and on PTB (with test set perplexity 56.6), with very limited computational resources (less than 10 GPU hours) for both tasks.
In cryo-electron microscopy (EM), molecular structures are determined from large numbers of projection images of individual particles. To harness the full power of this single-molecule information, we use the Bayesian inference of EM (BioEM) formalism. By ranking structural models using posterior probabilities calculated for individual images, BioEM in principle addresses the challenge of working with highly dynamic or heterogeneous systems not easily handled in traditional EM reconstruction. However, the calculation of these posteriors for large numbers of particles and models is computationally demanding. Here we present highly parallelized, GPU-accelerated computer software that performs this task efficiently. Our flexible formulation employs CUDA, OpenMP, and MPI parallelization combined with both CPU and GPU computing. The resulting BioEM software scales nearly ideally both on pure CPU and on CPU+GPU architectures, thus enabling Bayesian analysis of tens of thousands of images in a reasonable time. The general mathematical framework and robust algorithms are not limited to cryo-electron microscopy but can be generalized for electron tomography and other imaging experiments.
Convolutional neural networks have gained a remarkable success in computer vision. However, most usable network architectures are hand-crafted and usually require expertise and elaborate design. In this paper, we provide a block-wise network generation pipeline called BlockQNN which automatically builds high-performance networks using the Q-Learning paradigm with epsilon-greedy exploration strategy. The optimal network block is constructed by the learning agent which is trained to choose component layers sequentially. We stack the block to construct the whole auto-generated network. To accelerate the generation process, we also propose a distributed asynchronous framework and an early stop strategy. The block-wise generation brings unique advantages: (1) it yields state-of-the-art results in comparison to the hand-crafted networks on image classification, particularly, the best network generated by BlockQNN achieves 2.35% top-1 error rate on CIFAR-10. (2) it offers tremendous reduction of the search space in designing networks, spending only 3 days with 32 GPUs. A faster version can yield a comparable result with only 1 GPU in 20 hours. (3) it has strong generalizability in that the network built on CIFAR also performs well on the larger-scale dataset. The best network achieves very competitive accuracy of 82.0% top-1 and 96.0% top-5 on ImageNet.
Data parallelism has become a dominant method to scale Deep Neural Network (DNN) training across multiple nodes. Since synchronizing a large number of gradients of the local model can be a bottleneck for large-scale distributed training, compressing communication data has gained widespread attention recently. Among several recent proposed compression algorithms, Residual Gradient Compression (RGC) is one of the most successful approaches---it can significantly compress the transmitting message size (0.1\% of the gradient size) of each node and still achieve correct accuracy and the same convergence speed. However, the literature on compressing deep networks focuses almost exclusively on achieving good theoretical compression rate, while the efficiency of RGC in real distributed implementation has been less investigated. In this paper, we develop an RGC-based system that is able to reduce the end-to-end training time on real-world multi-GPU systems. Our proposed design called RedSync, which introduces a set of optimizations to reduce communication bandwidth requirement while introducing limited overhead. We evaluate the performance of RedSync on two different multiple GPU platforms, including 128 GPUs of a supercomputer and an 8-GPU server. Our test cases include image classification tasks on Cifar10 and ImageNet, and language modeling tasks on Penn Treebank and Wiki2 datasets. For DNNs featured with high communication to computation ratio, which have long been considered with poor scalability, RedSync brings significant performance improvements.
Chemical transport models (CTMs), which simulate air pollution transport, transformation, and removal, are computationally expensive, largely because of the computational intensity of the chemical mechanisms: systems of coupled differential equations representing atmospheric chemistry. Here we investigate the potential for machine learning to reproduce the behavior of a chemical mechanism, yet with reduced computational expense. We create a 17-layer residual multi-target regression neural network to emulate the Carbon Bond Mechanism Z (CBM-Z) gas-phase chemical mechanism. We train the network to match CBM-Z predictions of changes in concentrations of 77 chemical species after one hour, given a range of chemical and meteorological input conditions, which it is able to do with root-mean-square error (RMSE) of less than 1.97 ppb (median RMSE = 0.02 ppb), while achieving a 250x computational speedup. An additional 17x speedup (total 4250x speedup) is achieved by running the neural network on a graphics-processing unit (GPU). The neural network is able to reproduce the emergent behavior of the chemical system over diurnal cycles using Euler integration, but additional work is needed to constrain the propagation of errors as simulation time progresses.
Matrix factorization (MF) discovers latent features from observations, which has shown great promises in the fields of collaborative filtering, data compression, feature extraction, word embedding, etc. While many problem-specific optimization techniques have been proposed, alternating least square (ALS) remains popular due to its general applicability e.g. easy to handle positive-unlabeled inputs, fast convergence and parallelization capability. Current MF implementations are either optimized for a single machine or with a need of a large computer cluster but still are insufficient. This is because a single machine provides limited compute power for large-scale data while multiple machines suffer from the network communication bottleneck. To address the aforementioned challenge, accelerating ALS on graphics processing units (GPUs) is a promising direction. We propose the novel approach in enhancing the MF efficiency via both memory optimization and approximate computing. The former exploits GPU memory hierarchy to increase data reuse, while the later reduces unnecessary computing without hurting the convergence of learning algorithms. Extensive experiments on large-scale datasets show that our solution not only outperforms the competing CPU solutions by a large margin but also has a 2x-4x performance gain compared to the state-of-the-art GPU solutions. Our implementations are open-sourced and publicly available.
Detection of video shot transition is a crucial pre-processing step in video analysis. Previous studies are restricted on detecting sudden content changes between frames through similarity measurement and multi-scale operations are widely utilized to deal with transitions of various lengths. However, localization of gradual transitions are still under-explored due to the high visual similarity between adjacent frames. Cut shot transitions are abrupt semantic breaks while gradual shot transitions contain low-level spatial-temporal patterns caused by video effects in addition to the gradual semantic breaks, e.g. dissolve. In order to address the problem, we propose a structured network which is able to detect these two shot transitions using targeted models separately. Considering speed performance trade-offs, we design a smart framework. With one TITAN GPU, the proposed method can achieve a 30\(\times\) real-time speed. Experiments on public TRECVID07 and RAI databases show that our method outperforms the state-of-the-art methods. In order to train a high-performance shot transition detector, we contribute a new database ClipShots, which contains 128636 cut transitions and 38120 gradual transitions from 4039 online videos. ClipShots intentionally collect short videos for more hard cases caused by hand-held camera vibrations, large object motions, and occlusion.
Modern deep learning systems rely on (a) a hand-tuned neural network topology, (b) massive amounts of labeled training data, and (c) extensive training over large-scale compute resources to build a system that can perform efficient image classification or speech recognition. Unfortunately, we are still far away from implementing adaptive general purpose intelligent systems which would need to learn autonomously in unknown environments and may not have access to some or any of these three components. Reinforcement learning and evolutionary algorithm (EA) based methods circumvent this problem by continuously interacting with the environment and updating the models based on obtained rewards. However, deploying these algorithms on ubiquitous autonomous agents at the edge (robots/drones) demands extremely high energy-efficiency due to (i) tight power and energy budgets, (ii) continuous/lifelong interaction with the environment, (iii) intermittent or no connectivity to the cloud to run heavy-weight processing. To address this need, we present GENESYS, an HW-SW prototype of an EA-based learning system, that comprises a closed loop learning engine called EvE and an inference engine called ADAM. EvE can evolve the topology and weights of neural networks completely in hardware for the task at hand, without requiring hand-optimization or backpropagation training. ADAM continuously interacts with the environment and is optimized for efficiently running the irregular neural networks generated by EvE. GENESYS identifies and leverages multiple unique avenues of parallelism unique to EAs that we term 'gene'- level parallelism, and 'population'-level parallelism. We ran GENESYS with a suite of environments from OpenAI gym and observed 2-5 orders of magnitude higher energy-efficiency over state-of-the-art embedded and desktop CPU and GPU systems.
Object detection and classification in 3D is a key task in Automated Driving (AD). LiDAR sensors are employed to provide the 3D point cloud reconstruction of the surrounding environment, while the task of 3D object bounding box detection in real time remains a strong algorithmic challenge. In this paper, we build on the success of the one-shot regression meta-architecture in the 2D perspective image space and extend it to generate oriented 3D object bounding boxes from LiDAR point cloud. Our main contribution is in extending the loss function of YOLO v2 to include the yaw angle, the 3D box center in Cartesian coordinates and the height of the box as a direct regression problem. This formulation enables real-time performance, which is essential for automated driving. Our results are showing promising figures on KITTI benchmark, achieving real-time performance (40 fps) on Titan X GPU.
Deep Learning (DL) applications are gaining momentum in the realm of Artificial Intelligence, particularly after GPUs have demonstrated remarkable skills for accelerating their challenging computational requirements. Within this context, Convolutional Neural Network (CNN) models constitute a representative example of success on a wide set of complex applications, particularly on datasets where the target can be represented through a hierarchy of local features of increasing semantic complexity. In most of the real scenarios, the roadmap to improve results relies on CNN settings involving brute force computation, and researchers have lately proven Nvidia GPUs to be one of the best hardware counterparts for acceleration. Our work complements those findings with an energy study on critical parameters for the deployment of CNNs on flagship image and video applications: object recognition and people identification by gait, respectively. We evaluate energy consumption on four different networks based on the two most popular ones (ResNet/AlexNet): ResNet (167 layers), a 2D CNN (15 layers), a CaffeNet (25 layers) and a ResNetIm (94 layers) using batch sizes of 64, 128 and 256, and then correlate those with speed-up and accuracy to determine optimal settings. Experimental results on a multi-GPU server endowed with twin Maxwell and twin Pascal Titan X GPUs demonstrate that energy correlates with performance and that Pascal may have up to 40% gains versus Maxwell. Larger batch sizes extend performance gains and energy savings, but we have to keep an eye on accuracy, which sometimes shows a preference for small batches. We expect this work to provide a preliminary guidance for a wide set of CNN and DL applications in modern HPC times, where the GFLOPS/w ratio constitutes the primary goal.
Convolutional neural networks have recently achieved significant breakthroughs in various image classification tasks. However, they are computationally expensive,which can make their feasible mplementation on embedded and low-power devices difficult. In this paper convolutional neural network binarization is implemented on GPU-based platforms for real-time inference on resource constrained devices. In binarized networks, all weights and intermediate computations between layers are quantized to +1 and -1, allowing multiplications and additions to be replaced with bit-wise operations between 32-bit words. This representation completely eliminates the need for floating point multiplications and additions and decreases both the computational load and the memory footprint compared to a full-precision network implemented in floating point, making it well-suited for resource-constrained environments. We compare the performance of our implementation with an equivalent floating point implementation on one desktop and two embedded GPU platforms. Our implementation achieves a maximum speed up of 7. 4X with only 4.4% loss in accuracy compared to a reference implementation.
Deep Neural Networks(DNNs) require huge GPU memory when training on modern image/video databases. Unfortunately, the GPU memory is physically finite, which limits the image resolutions and batch sizes that could be used in training for better DNN performance. Unlike solutions that require physically upgrade GPUs, the Gradient CheckPointing(GCP) training trades computation for more memory beyond existing GPU hardware. GCP only stores a subset of intermediate tensors, called Gradient Checkpoints (GCs), during forward. Then during backward, extra local forwards are conducted to compute the missing tensors. The total training memory cost becomes the sum of (1) the memory cost of the gradient checkpoints and (2) the maximum memory cost of local forwards. To achieve maximal memory cut-offs, one needs optimal algorithms to select GCs. Existing GCP approaches rely on either manual input of GCs or heuristics-based GC search on Linear Computation Graphs (LCGs), and cannot apply to Arbitrary Computation Graphs(ACGs). In this paper, we present theories and optimal algorithms on GC selection that, for the first time, are applicable to ACGs and achieve the maximal memory cut-offs. Extensive experiments show that our approach not only outperforms existing approaches (only applicable on LCGs), and is applicable to a vast family of LCG and ACG networks, such as Alexnet, VGG, ResNet, Densenet, Inception Net and highly complicated DNNs by Network Architecture Search. Our work enables GCP training on ACGs, and cuts off up-to 80% of training memory with a moderate time overhead (~30%-50%). Codes are available
Convolutional Neural Networks (CNN) for image recognition tasks are seeing rapid advances in the available architectures and how networks are trained based on large computational infrastructure and standard datasets with millions of images. In contrast, performance and time constraints for example, of small devices and free cloud GPUs necessitate efficient network training (i.e., highest accuracy in the shortest inference time possible), often on small datasets. Here, we hypothesize that initially decreasing image size during training makes the training process more efficient, because pre-shaping weights with small images and later utilizing these weights with larger images reduces initial network parameters and total inference time. We test this Efficient Network TRaining (ENTR) Hypothesis by training pre-trained Residual Network (ResNet) models (ResNet18, 34, & 50) on three small datasets (steel microstructures, bee images, and geographic aerial images) with a free cloud GPU. Based on three training regimes of i) not, ii) gradually or iii) in one step increasing image size over the training process, we show that initially reducing image size increases training efficiency consistently across datasets and networks. We interpret these results mechanistically in the framework of regularization theory. Support for the ENTR hypothesis is an important contribution, because network efficiency improvements for image recognition tasks are needed for practical applications. In the future, it will be exciting to see how the ENTR hypothesis holds for large standard datasets like ImageNet or CIFAR, to better understand the underlying mechanisms, and how these results compare to other fields such as structural learning.
Modern datasets and models are notoriously difficult to explore and analyze due to their inherent high dimensionality and massive numbers of samples. Existing visualization methods which employ dimensionality reduction to two or three dimensions are often inefficient and/or ineffective for these datasets. This paper introduces t-SNE-CUDA, a GPU-accelerated implementation of t-distributed Symmetric Neighbor Embedding (t-SNE) for visualizing datasets and models. t-SNE-CUDA significantly outperforms current implementations with 50-700x speedups on the CIFAR-10 and MNIST datasets. These speedups enable, for the first time, visualization of the neural network activations on the entire ImageNet dataset - a feat that was previously computationally intractable. We also demonstrate visualization performance in the NLP domain by visualizing the GloVe embedding vectors. From these visualizations, we can draw interesting conclusions about using the L2 metric in these embedding spaces. t-SNE-CUDA is publicly available athttps://github.com/CannyLab/tsne-cuda
The combination of spiking neural networks and event-based vision sensors holds the potential of highly efficient and high-bandwidth optical flow estimation. This paper presents the first hierarchical spiking architecture in which motion (direction and speed) selectivity emerges in an unsupervised fashion from the raw stimuli generated with an event-based camera. A novel adaptive neuron model and stable spike-timing-dependent plasticity formulation are at the core of this neural network governing its spike-based processing and learning, respectively. After convergence, the neural architecture exhibits the main properties of biological visual motion systems, namely feature extraction and local and global motion perception. Convolutional layers with input synapses characterized by single and multiple transmission delays are employed for feature and local motion perception, respectively; while global motion selectivity emerges in a final fully-connected layer. The proposed solution is validated using synthetic and real event sequences. Along with this paper, we provide the cuSNN library, a framework that enables GPU-accelerated simulations of large-scale spiking neural networks. Source code and samples are available at https://github.com/tudelft/cuSNN.
Synchronized stochastic gradient descent (SGD) optimizers with data parallelism are widely used in training large-scale deep neural networks. Although using larger mini-batch sizes can improve the system scalability by reducing the communication-to-computation ratio, it may hurt the generalization ability of the models. To this end, we build a highly scalable deep learning training system for dense GPU clusters with three main contributions: (1) We propose a mixed-precision training method that significantly improves the training throughput of a single GPU without losing accuracy. (2) We propose an optimization approach for extremely large mini-batch size (up to 64k) that can train CNN models on the ImageNet dataset without losing accuracy. (3) We propose highly optimized all-reduce algorithms that achieve up to 3x and 11x speedup on AlexNet and ResNet-50 respectively than NCCL-based training on a cluster with 1024 Tesla P40 GPUs. On training ResNet-50 with 90 epochs, the state-of-the-art GPU-based system with 1024 Tesla P100 GPUs spent 15 minutes and achieved 74.9\% top-1 test accuracy, and another KNL-based system with 2048 Intel KNLs spent 20 minutes and achieved 75.4\% accuracy. Our training system can achieve 75.8\% top-1 test accuracy in only 6.6 minutes using 2048 Tesla P40 GPUs. When training AlexNet with 95 epochs, our system can achieve 58.7\% top-1 test accuracy within 4 minutes, which also outperforms all other existing systems.
Weight pruning methods of DNNs have been demonstrated to achieve a good model pruning rate without loss of accuracy, thereby alleviating the significant computation/storage requirements of large-scale DNNs. Structured weight pruning methods have been proposed to overcome the limitation of irregular network structure and demonstrated actual GPU acceleration. However, in prior work the pruning rate (degree of sparsity) and GPU acceleration are limited (to less than 50%) when accuracy needs to be maintained. In this work,we overcome these limitations by proposing a unified, systematic framework of structured weight pruning for DNNs. It is a framework that can be used to induce different types of structured sparsity, such as filter-wise, channel-wise, and shape-wise sparsity, as well non-structured sparsity. The proposed framework incorporates stochastic gradient descent with ADMM, and can be understood as a dynamic regularization method in which the regularization target is analytically updated in each iteration. Without loss of accuracy on the AlexNet model, we achieve 2.58X and 3.65X average measured speedup on two GPUs, clearly outperforming the prior work. The average speedups reach 3.15X and 8.52X when allowing a moderate ac-curacy loss of 2%. In this case the model compression for convolutional layers is 15.0X, corresponding to 11.93X measured CPU speedup. Our experiments on ResNet model and on other data sets like UCF101 and CIFAR-10 demonstrate the consistently higher performance of our framework.
We present a novel learning-based framework for face reenactment. The proposed method, known as ReenactGAN, is capable of transferring facial movements and expressions from monocular video input of an arbitrary person to a target person. Instead of performing a direct transfer in the pixel space, which could result in structural artifacts, we first map the source face onto a boundary latent space. A transformer is subsequently used to adapt the boundary of source face to the boundary of target face. Finally, a target-specific decoder is used to generate the reenacted target face. Thanks to the effective and reliable boundary-based transfer, our method can perform photo-realistic face reenactment. In addition, ReenactGAN is appealing in that the whole reenactment process is purely feed-forward, and thus the reenactment process can run in real-time (30 FPS on one GTX 1080 GPU). Dataset and model will be publicly available at https://wywu.github.io/projects/ReenactGAN/ReenactGAN.html
Motivation: In this paper we present the latest release of EBIC, a next-generation biclustering algorithm for mining genetic data. The major contribution of this paper is adding support for big data, making it possible to efficiently run large genomic data mining analyses. Additional enhancements include integration with R and Bioconductor and an option to remove influence of missing value on the final result. Results: EBIC was applied to datasets of different sizes, including a large DNA methylation dataset with 436,444 rows. For the largest dataset we observed over 6.6 fold speedup in computation time on a cluster of 8 GPUs compared to running the method on a single GPU. This proves high scalability of the algorithm. Availability: The latest version of EBIC could be downloaded from http://github.com/EpistasisLab/ebic . Installation and usage instructions are also available online.
This paper presents Tofu, a system that partitions very large DNN models across multiple GPU devices to reduce per-GPU memory footprint. Tofu is designed to partition a dataflow graph of fine-grained tensor operators in order to work transparently with a general-purpose deep learning platform like MXNet. In order to automatically partition each operator, we propose to describe the semantics of an operator in a simple language which represents tensors as lambda functions mapping from tensor coordinates to values. To optimally partition different operators in a dataflow graph, Tofu uses a recursive search algorithm that minimizes the total communication cost. Our experiments on an 8-GPU machine show that Tofu enables the training of very large CNN and RNN models. It also achieves 25% - 400% speedup over alternative approaches to train very large models.