We have developed convolutional neural networks (CNN) for a facial expression recognition task. The goal is to classify each facial image into one of the seven facial emotion categories considered in this study. We trained CNN models with different depth using gray-scale images. We developed our models in Torch and exploited Graphics Processing Unit (GPU) computation in order to expedite the training process. In addition to the networks performing based on raw pixel data, we employed a hybrid feature strategy by which we trained a novel CNN model with the combination of raw pixel data and Histogram of Oriented Gradients (HOG) features. To reduce the overfitting of the models, we utilized different techniques including dropout and batch normalization in addition to L2 regularization. We applied cross validation to determine the optimal hyper-parameters and evaluated the performance of the developed models by looking at their training histories. We also present the visualization of different layers of a network to show what features of a face can be learned by CNN models.
We focus on the challenging task of real-time semantic segmentation in this paper. It finds many practical applications and yet is with fundamental difficulty of reducing a large portion of computation for pixel-wise label inference. We propose an image cascade network (ICNet) that incorporates multi-resolution branches under proper label guidance to address this challenge. We provide in-depth analysis of our framework and introduce the cascade feature fusion unit to quickly achieve high-quality segmentation. Our system yields real-time inference on a single GPU card with decent quality results evaluated on challenging datasets like Cityscapes, CamVid and COCO-Stuff.
Text to image transformation for input to neural networks requires intermediate steps. This paper attempts to present a new approach to pixel normalization so as to convert textual data into image, suitable as input for neural networks. This method can be further improved by its Graphics Processing Unit (GPU) implementation to provide significant speedup in computational time.
Popular deep learning frameworks require users to fine-tune their memory usage so that the training data of a deep neural network (DNN) fits within the GPU physical memory. Prior work tries to address this restriction by virtualizing the memory usage of DNNs, enabling both CPU and GPU memory to be utilized for memory allocations. Despite its merits, virtualizing memory can incur significant performance overheads when the time needed to copy data back and forth from CPU memory is higher than the latency to perform the computations required for DNN forward and backward propagation. We introduce a high-performance virtualization strategy based on a "compressing DMA engine" (cDMA) that drastically reduces the size of the data structures that are targeted for CPU-side allocations. The cDMA engine offers an average 2.6x (maximum 13.8x) compression ratio by exploiting the sparsity inherent in offloaded data, improving the performance of virtualized DNNs by an average 32% (maximum 61%).
The Ant Colony System (ACS) is, next to Ant Colony Optimization (ACO) and the MAX-MIN Ant System (MMAS), one of the most efficient metaheuristic algorithms inspired by the behavior of ants. In this article we present three novel parallel versions of the ACS for the graphics processing units (GPUs). To the best of our knowledge, this is the first such work on the ACS which shares many key elements of the ACO and the MMAS, but differences in the process of building solutions and updating the pheromone trails make obtaining an efficient parallel version for the GPUs a difficult task. The proposed parallel versions of the ACS differ mainly in their implementations of the pheromone memory. The first two use the standard pheromone matrix, and the third uses a novel selective pheromone memory. Computational experiments conducted on several Travelling Salesman Problem (TSP) instances of sizes ranging from 198 to 2392 cities showed that the parallel ACS on Nvidia Kepler GK104 GPU (1536 CUDA cores) is able to obtain a speedup up to 24.29x vs the sequential ACS running on a single core of Intel Xeon E5-2670 CPU. The parallel ACS with the selective pheromone memory achieved speedups up to 16.85x, but in most cases the obtained solutions were of significantly better quality than for the sequential ACS.
We investigate GPU-based parallelization of Iterative-Deepening A* (IDA*). We show that straightforward thread-based parallelization techniques which were previously proposed for massively parallel SIMD processors perform poorly due to warp divergence and load imbalance. We propose Block-Parallel IDA* (BPIDA*), which assigns the search of a subtree to a block (a group of threads with access to fast shared memory) rather than a thread. On the 15-puzzle, BPIDA* on a NVIDIA GRID K520 with 1536 CUDA cores achieves a speedup of 4.98 compared to a highly optimized sequential IDA* implementation on a Xeon E5-2670 core.
In recent years deep learning algorithms have shown extremely high performance on machine learning tasks such as image classification and speech recognition. In support of such applications, various FPGA accelerator architectures have been proposed for convolutional neural networks (CNNs) that enable high performance for classification tasks at lower power than CPU and GPU processors. However, to date, there has been little research on the use of FPGA implementations of deconvolutional neural networks (DCNNs). DCNNs, also known as generative CNNs, encode high-dimensional probability distributions and have been widely used for computer vision applications such as scene completion, scene segmentation, image creation, image denoising, and super-resolution imaging. We propose an FPGA architecture for deconvolutional networks built around an accelerator which effectively handles the complex memory access patterns needed to perform strided deconvolutions, and that supports convolution as well. We also develop a three-step design optimization method that systematically exploits statistical analysis, design space exploration and VLSI optimization. To verify our FPGA deconvolutional accelerator design methodology we train DCNNs offline on two representative datasets using the generative adversarial network method (GAN) run on Tensorflow, and then map these DCNNs to an FPGA DCNN-plus-accelerator implementation to perform generative inference on a Xilinx Zynq-7000 FPGA. Our DCNN implementation achieves a peak performance density of 0.012 GOPs/DSP.
We present a machine learning-based approach to lossy image compression which outperforms all existing codecs, while running in real-time. Our algorithm typically produces files 2.5 times smaller than JPEG and JPEG 2000, 2 times smaller than WebP, and 1.7 times smaller than BPG on datasets of generic images across all quality levels. At the same time, our codec is designed to be lightweight and deployable: for example, it can encode or decode the Kodak dataset in around 10ms per image on GPU. Our architecture is an autoencoder featuring pyramidal analysis, an adaptive coding module, and regularization of the expected codelength. We also supplement our approach with adversarial training specialized towards use in a compression setting: this enables us to produce visually pleasing reconstructions for very low bitrates.
We propose a novel framework for efficient parallelization of deep reinforcement learning algorithms, enabling these algorithms to learn from multiple actors on a single machine. The framework is algorithm agnostic and can be applied to on-policy, off-policy, value based and policy gradient based algorithms. Given its inherent parallelism, the framework can be efficiently implemented on a GPU, allowing the usage of powerful models while significantly reducing training time. We demonstrate the effectiveness of our framework by implementing an advantage actor-critic algorithm on a GPU, using on-policy experiences and employing synchronous updates. Our algorithm achieves state-of-the-art performance on the Atari domain after only a few hours of training. Our framework thus opens the door for much faster experimentation on demanding problem domains. Our implementation is open-source and is made public at https://github.com/alfredvc/paac
Deep Neural Networks (DNN) achieve human level performance in many image analytics tasks but DNNs are mostly deployed to GPU platforms that consume a considerable amount of power. New hardware platforms using lower precision arithmetic achieve drastic reductions in power consumption. More recently, brain-inspired spiking neuromorphic chips have achieved even lower power consumption, on the order of milliwatts, while still offering real-time processing. However, for deploying DNNs to energy efficient neuromorphic chips the incompatibility between continuous neurons and synaptic weights of traditional DNNs, discrete spiking neurons and synapses of neuromorphic chips need to be overcome. Previous work has achieved this by training a network to learn continuous probabilities, before it is deployed to a neuromorphic architecture, such as IBM TrueNorth Neurosynaptic System, by random sampling these probabilities. The main contribution of this paper is a new learning algorithm that learns a TrueNorth configuration ready for deployment. We achieve this by training directly a binary hardware crossbar that accommodates the TrueNorth axon configuration constrains and we propose a different neuron model. Results of our approach trained on electroencephalogram (EEG) data show a significant improvement with previous work (76% vs 86% accuracy) while maintaining state of the art performance on the MNIST handwritten data set.
Anatomical and biophysical modeling of left atrium (LA) and proximal pulmonary veins (PPVs) is important for clinical management of several cardiac diseases. Magnetic resonance imaging (MRI) allows qualitative assessment of LA and PPVs through visualization. However, there is a strong need for an advanced image segmentation method to be applied to cardiac MRI for quantitative analysis of LA and PPVs. In this study, we address this unmet clinical need by exploring a new deep learning-based segmentation strategy for quantification of LA and PPVs with high accuracy and heightened efficiency. Our approach is based on a multi-view convolutional neural network (CNN) with an adaptive fusion strategy and a new loss function that allows fast and more accurate convergence of the backpropagation based optimization. After training our network from scratch by using more than 60K 2D MRI images (slices), we have evaluated our segmentation strategy to the STACOM 2013 cardiac segmentation challenge benchmark. Qualitative and quantitative evaluations, obtained from the segmentation challenge, indicate that the proposed method achieved the state-of-the-art sensitivity (90%), specificity (99%), precision (94%), and efficiency levels (10 seconds in GPU, and 7.5 minutes in CPU).
Three approaches to implement genetic programming on GPU hardware are compilation, interpretation and direct generation of machine code. The compiled approach is known to have a prohibitive overhead compared to other two. This paper investigates methods to accelerate compilation of individuals for genetic programming on GPU hardware. We apply in-process compilation to minimize the compilation overhead at each generation; and we investigate ways to parallelize in-process compilation. In-process compilation doesn't lend itself to trivial parallelization with threads; we propose a multiprocess parallelization using memory sharing and operating systems interprocess communication primitives. With parallelized compilation we achieve further reductions on compilation overhead. Another contribution of this work is the code framework we built in C# for the experiments. The framework makes it possible to build arbitrary grammatical genetic programming experiments that run on GPU with minimal extra coding effort, and is available as open source.
We propose a partially learned approach for the solution of ill posed inverse problems with not necessarily linear forward operators. The method builds on ideas from classical regularization theory and recent advances in deep learning to perform learning while making use of prior information about the inverse problem encoded in the forward operator, noise model and a regularizing functional. The method results in a gradient-like iterative scheme, where the "gradient" component is learned using a convolutional network that includes the gradients of the data discrepancy and regularizer as input in each iteration. We present results of such a partially learned gradient scheme on a non-linear tomographic inversion problem with simulated data from both the Sheep-Logan phantom as well as a head CT. The outcome is compared against FBP and TV reconstruction and the proposed method provides a 5.4 dB PSNR improvement over the TV reconstruction while being significantly faster, giving reconstructions of 512 x 512 volumes in about 0.4 seconds using a single GPU.
Based on Alan Turing's proposition on AI and computing machinery, which shaped Computing as we know it today, the new AI computing machinery should comprise a universal computer and a universal learning machine. The later should understand linear algebra natively to overcome the slowdown of Moore's law. In such a universal learnig machine, a computing unit does not need to keep the legacy of a universal computing core. The data can be distributed to the computing units, and the results can be collected from them through Collective Streaming, reminiscent of Collective Communication in Supercomputing. It is not necessary to use a GPU-like deep memory hierarchy, nor a TPU-like fine-grain mesh.
We propose an object detection method that improves the accuracy of the conventional SSD (Single Shot Multibox Detector), which is one of the top object detection algorithms in both aspects of accuracy and speed. The performance of a deep network is known to be improved as the number of feature maps increases. However, it is difficult to improve the performance by simply raising the number of feature maps. In this paper, we propose and analyze how to use feature maps effectively to improve the performance of the conventional SSD. The enhanced performance was obtained by changing the structure close to the classifier network, rather than growing layers close to the input data, e.g., by replacing VGGNet with ResNet. The proposed network is suitable for sharing the weights in the classifier networks, by which property, the training can be faster with better generalization power. For the Pascal VOC 2007 test set trained with VOC 2007 and VOC 2012 training sets, the proposed network with the input size of 300 x 300 achieved 78.5% mAP (mean average precision) at the speed of 35.0 FPS (frame per second), while the network with a 512 x 512 sized input achieved 80.8% mAP at 16.6 FPS using Nvidia Titan X GPU. The proposed network shows state-of-the-art mAP, which is better than those of the conventional SSD, YOLO, Faster-RCNN and RFCN. Also, it is faster than Faster-RCNN and RFCN.
Binary Neural Networks (BNNs) can drastically reduce memory size and accesses by applying bit-wise operations instead of standard arithmetic operations. Therefore it could significantly improve the efficiency and lower the energy consumption at runtime, which enables the application of state-of-the-art deep learning models on low power devices. BMXNet is an open-source BNN library based on MXNet, which supports both XNOR-Networks and Quantized Neural Networks. The developed BNN layers can be seamlessly applied with other standard library components and work in both GPU and CPU mode. BMXNet is maintained and developed by the multimedia research group at Hasso Plattner Institute and released under Apache license. Extensive experiments validate the efficiency and effectiveness of our implementation. The BMXNet library, several sample projects, and a collection of pre-trained binary deep models are available for download at https://github.com/hpi-xnor
End-to-end training of automated speech recognition (ASR) systems requires massive data and compute resources. We explore transfer learning based on model adaptation as an approach for training ASR models under constrained GPU memory, throughput and training data. We conduct several systematic experiments adapting a Wav2Letter convolutional neural network originally trained for English ASR to the German language. We show that this technique allows faster training on consumer-grade resources while requiring less training data in order to achieve the same accuracy, thereby lowering the cost of training ASR models in other languages. Model introspection revealed that small adaptations to the network's weights were sufficient for good performance, especially for inner layers.
Multi-start algorithms are a common and effective tool for metaheuristic searches. In this paper we amplify multi-start capabilities by employing the parallel processing power of the graphics processer unit (GPU) to quickly generate a diverse starting set of solutions for the Unconstrained Binary Quadratic Optimization Problem which are evaluated and used to implement screening methods to select solutions for further optimization. This method is implemented as an initial high quality solution generation phase prior to a secondary steepest ascent search and a comparison of results to best known approaches on benchmark unconstrained binary quadratic problems demonstrates that GPU-enabled diversified multi-start with screening quickly yields very good results.
We present a simple and effective architecture for fine-grained visual recognition called Bilinear Convolutional Neural Networks (B-CNNs). These networks represent an image as a pooled outer product of features derived from two CNNs and capture localized feature interactions in a translationally invariant manner. B-CNNs belong to the class of orderless texture representations but unlike prior work they can be trained in an end-to-end manner. Our most accurate model obtains 84.1%, 79.4%, 86.9% and 91.3% per-image accuracy on the Caltech-UCSD birds [67], NABirds [64], FGVC aircraft [42], and Stanford cars [33] dataset respectively and runs at 30 frames-per-second on a NVIDIA Titan X GPU. We then present a systematic analysis of these networks and show that (1) the bilinear features are highly redundant and can be reduced by an order of magnitude in size without significant loss in accuracy, (2) are also effective for other image classification tasks such as texture and scene recognition, and (3) can be trained from scratch on the ImageNet dataset offering consistent improvements over the baseline architecture. Finally, we present visualizations of these models on various datasets using top activations of neural units and gradient-based inversion techniques. The source code for the complete system is available at http://vis-www.cs.umass.edu/bcnn.
In this paper, we explore optimizations to run Recurrent Neural Network (RNN) models locally on mobile devices. RNN models are widely used for Natural Language Processing, Machine Translation, and other tasks. However, existing mobile applications that use RNN models do so on the cloud. To address privacy and efficiency concerns, we show how RNN models can be run locally on mobile devices. Existing work on porting deep learning models to mobile devices focus on Convolution Neural Networks (CNNs) and cannot be applied directly to RNN models. In response, we present MobiRNN, a mobile-specific optimization framework that implements GPU offloading specifically for mobile GPUs. Evaluations using an RNN model for activity recognition shows that MobiRNN does significantly decrease the latency of running RNN models on phones.