Convolutional neural network (CNN) architectures have traditionally been explored by human experts in a manual search process that is time-consuming and ineffectively explores the massive space of potential solutions. Neural architecture search (NAS) methods automatically search the space of neural network hyperparameters in order to find optimal task-specific architectures. NAS methods have discovered CNN architectures that achieve state-of-the-art performance in image classification among other tasks, however the application of NAS to image-to-image regression problems such as image restoration is sparse. This paper proposes a NAS method that performs computationally efficient evolutionary search of a minimally constrained network architecture search space. The performance of architectures discovered by the proposed method is evaluated on a variety of image restoration tasks applied to the ImageNet64x64 dataset, and compared with human-engineered CNN architectures. The best neural architectures discovered using only 2 GPU-hours of evolutionary search exhibit comparable performance to the human-engineered baseline architecture.
Purpose: To perform and evaluate water-fat signal separation of whole-body gradient echo scans using convolutional neural networks. Methods: Whole-body gradient echo scans of 240 subjects, each consisting of 5 bipolar echoes, were used. Reference fat fraction maps were created using a conventional method. Convolutional neural networks, more specifically 2D U-nets, were trained using 5-fold cross-validation with 1 or several echoes as input, using the squared difference between the output and the reference fat fraction maps as the loss function. The outputs of the networks were assessed by the loss function, measured liver fat fractions, and visually. Training was performed using a graphics processing unit (GPU). Inference was performed using the GPU as well as a central processing unit (CPU). Results: The loss curves indicated convergence, and the final loss of the validation data decreased when using more echoes as input. The liver fat fractions could be estimated using only 1 echo, but results were improved by use of more echoes. Visual assessment found the quality of the outputs of the networks to be similar to the reference even when using only 1 echo, with slight improvements when using more echoes. Training a network took at most 28.6 h. Inference time of a whole-body scan took at most 3.7 s using the GPU and 5.8 min using the CPU. Conclusion: It is possible to perform water-fat signal separation of whole-body gradient echo scans using convolutional neural networks. Separation was possible using only 1 echo, although using more echoes improved the results.
Recent advancements in parallel computing, GPU technology and deep learning provide a new platform for complex image processing tasks such as person detection to flourish. Person detection is fundamental preliminary operation for several high level computer vision tasks. One industry that can significantly benefit from person detection is retail. In recent years, various studies attempt to find an optimal solution for person detection using neural networks and deep learning. This study conducts a comparison among the state of the art deep learning base object detector with the focus on person detection performance in indoor environments. Performance of various implementations of YOLO, SSD, RCNN, R-FCN and SqueezeDet have been assessed using our in-house proprietary dataset which consists of over 10 thousands indoor images captured form shopping malls, retails and stores. Experimental results indicate that, Tiny YOLO-416 and SSD (VGG-300) are the fastest and Faster-RCNN (Inception ResNet-v2) and R-FCN (ResNet-101) are the most accurate detectors investigated in this study. Further analysis shows that YOLO v3-416 delivers relatively accurate result in a reasonable amount of time, which makes it an ideal model for person detection in embedded platforms.
Designing accurate and efficient ConvNets for mobile devices is challenging because the design space is combinatorially large. Due to this, previous neural architecture search (NAS) methods are computationally expensive. ConvNet architecture optimality depends on factors such as input resolution and target devices. However, existing approaches are too expensive for case-by-case redesigns. Also, previous work focuses primarily on reducing FLOPs, but FLOP count does not always reflect actual latency. To address these, we propose a differentiable neural architecture search (DNAS) framework that uses gradient-based methods to optimize ConvNet architectures, avoiding enumerating and training individual architectures separately as in previous methods. FBNets, a family of models discovered by DNAS surpass state-of-the-art models both designed manually and generated automatically. FBNet-B achieves 74.1% top-1 accuracy on ImageNet with 295M FLOPs and 23.1 ms latency on a Samsung S8 phone, 2.4x smaller and 1.5x faster than MobileNetV2-1.3 with similar accuracy. Despite higher accuracy and lower latency than MnasNet, we estimate FBNet-B's search cost is 420x smaller than MnasNet's, at only 216 GPU-hours. Searched for different resolutions and channel sizes, FBNets achieve 1.5% to 6.4% higher accuracy than MobileNetV2. The smallest FBNet achieves 50.2% accuracy and 2.9 ms latency (345 frames per second) on a Samsung S8. Over a Samsung-optimized FBNet, the iPhone-X-optimized model achieves a 1.4x speedup on an iPhone X.
In this paper we proposed an ultimate theory to solve the multi-target control problem through its introduction to the machine learning framework in automatic driving, which explored the implementation of excellent drivers' knowledge acquisition. Nowadays there exist some core problems that have not been fully realized by the researchers in automatic driving, such as the optimal way to control the multi-target objective functions of energy saving, safe driving, headway distance control and comfort driving, as well as the resolvability of the networks that automatic driving relied on and the high-performance chips like GPU on the complex driving environments. According to these problems, we developed a new theory to map multitarget objective functions in different spaces into the same one and thus introduced a machine learning framework of SDL(Super Deep Learning) for optimal multi-targetcontrol based on knowledge acquisition. We will present in this paper the optimal multi-target control by combining the fuzzy relationship of each multi-target objective function and the implementation of excellent drivers' knowledge acquired by machine learning. Theoretically, the impact of this method will exceed that of the fuzzy control method used in automatic train.
The Earth Mover's Distance (EMD) is a state-of-the art metric for comparing discrete probability distributions, but its high distinguishability comes at a high cost in computational complexity. Even though linear-complexity approximation algorithms have been proposed to improve its scalability, these algorithms are either limited to vector spaces with only a few dimensions or they become ineffective when the degree of overlap between the probability distributions is high. We propose novel approximation algorithms that overcome both of these limitations, yet still achieve linear time complexity. All our algorithms are data parallel, and thus, we take advantage of massively parallel computing engines, such as Graphics Processing Units (GPUs). On the popular text-based 20 Newsgroups dataset, the new algorithms are four orders of magnitude faster than a multi-threaded CPU implementation of Word Mover's Distance and match its nearest-neighbors-search accuracy. On MNIST images, the new algorithms are four orders of magnitude faster than a GPU implementation of the Sinkhorn's algorithm while offering a slightly higher nearest-neighbors-search accuracy.
Using Intel's Loihi neuromorphic research chip and ABR's Nengo Deep Learning toolkit, we analyze the inference speed, dynamic power consumption, and energy cost per inference of a two-layer neural network keyword spotter trained to recognize a single phrase. We perform comparative analyses of this keyword spotter running on more conventional hardware devices including a CPU, a GPU, Nvidia's Jetson TX1, and the Movidius Neural Compute Stick. Our results indicate that for this inference application, Loihi outperforms all of these alternatives on an energy cost per inference basis while maintaining equivalent inference accuracy. Furthermore, an analysis of tradeoffs between network size, inference speed, and energy cost indicates that Loihi's comparative advantage over other low-power computing devices improves for larger networks.
Neural architecture search (NAS) has a great impact by automatically designing effective neural network architectures. However, the prohibitive computational demand of conventional NAS algorithms (e.g. $10^4$ GPU hours) makes it difficult to \emph{directly} search the architectures on large-scale tasks (e.g. ImageNet). Differentiable NAS can reduce the cost of GPU hours via a continuous representation of network architecture but suffers from the high GPU memory consumption issue (grow linearly w.r.t. candidate set size). As a result, they need to utilize~\emph{proxy} tasks, such as training on a smaller dataset, or learning with only a few blocks, or training just for a few epochs. These architectures optimized on proxy tasks are not guaranteed to be optimal on the target task. In this paper, we present \emph{ProxylessNAS} that can \emph{directly} learn the architectures for large-scale target tasks and target hardware platforms. We address the high memory consumption issue of differentiable NAS and reduce the computational cost (GPU hours and GPU memory) to the same level of regular training while still allowing a large candidate set. Experiments on CIFAR-10 and ImageNet demonstrate the effectiveness of directness and specialization. On CIFAR-10, our model achieves 2.08\% test error with only 5.7M parameters, better than the previous state-of-the-art architecture AmoebaNet-B, while using 6$\times$ fewer parameters. On ImageNet, our model achieves 3.1\% better top-1 accuracy than MobileNetV2, while being 1.2$\times$ faster with measured GPU latency. We also apply ProxylessNAS to specialize neural architectures for hardware with direct hardware metrics (e.g. latency) and provide insights for efficient CNN architecture design.
Speech, Music and Noise classification/segmentation is an important preprocessing step for audio processing/indexing. To this end, we propose a novel 1D Convolutional Neural Network (CNN) - SwishNet. It is a fast and lightweight architecture that operates on MFCC features which is suitable to be added to the front-end of an audio processing pipeline. We showed that the performance of our network can be improved by distilling knowledge from a 2D CNN, pretrained on ImageNet. We investigated the performance of our network on the MUSAN corpus - an openly available comprehensive collection of noise, music and speech samples, suitable for deep learning. The proposed network achieved high overall accuracy in clip (length of 0.5-2s) classification (>97% accuracy) and frame-wise segmentation (>93% accuracy) tasks with even higher accuracy (>99%) in speech/non-speech discrimination task. To verify the robustness of our model, we trained it on MUSAN and evaluated it on a different corpus - GTZAN and found good accuracy with very little fine-tuning. We also demonstrated that our model is fast on both CPU and GPU, consumes a low amount of memory and is suitable for implementation in embedded systems.
Determining the position and orientation of a calibrated camera from a single image with respect to a 3D model is an essential task for many applications. When 2D-3D correspondences can be obtained reliably, perspective-n-point solvers can be used to recover the camera pose. However, without the pose it is non-trivial to find cross-modality correspondences between 2D images and 3D models, particularly when the latter only contains geometric information. Consequently, the problem becomes one of estimating pose and correspondences jointly. Since outliers and local optima are so prevalent, robust objective functions and global search strategies are desirable. Hence, we cast the problem as a 2D-3D mixture model alignment task and propose the first globally-optimal solution to this formulation under the robust $L_2$ distance between mixture distributions. We search the 6D camera pose space using branch-and-bound, which requires novel bounds, to obviate the need for a pose estimate and guarantee global optimality. To accelerate convergence, we integrate local optimization, implement GPU bound computations, and provide an intuitive way to incorporate side information such as semantic labels. The algorithm is evaluated on challenging synthetic and real datasets, outperforming existing approaches and reliably converging to the global optimum.
This paper describes AutoFocus, an efficient multi-scale inference algorithm for deep-learning based object detectors. Instead of processing an entire image pyramid, AutoFocus adopts a coarse to fine approach and only processes regions which are likely to contain small objects at finer scales. This is achieved by predicting category agnostic segmentation maps for small objects at coarser scales, called FocusPixels. FocusPixels can be predicted with high recall, and in many cases, they only cover a small fraction of the entire image. To make efficient use of FocusPixels, an algorithm is proposed which generates compact rectangular FocusChips which enclose FocusPixels. The detector is only applied inside FocusChips, which reduces computation while processing finer scales. Different types of error can arise when detections from FocusChips of multiple scales are combined, hence techniques to correct them are proposed. AutoFocus obtains an mAP of 47.9% (68.3% at 50% overlap) on the COCO test-dev set while processing 6.4 images per second on a Titan X (Pascal) GPU. This is 2.5X faster than our multi-scale baseline detector and matches its mAP. The number of pixels processed in the pyramid can be reduced by 5X with a 1% drop in mAP. AutoFocus obtains more than 10% mAP gain compared to RetinaNet but runs at the same speed with the same ResNet-101 backbone.
We present an online approach to efficiently and simultaneously detect and track the 2D pose of multiple people in a video sequence. We build upon Part Affinity Field (PAF) representation designed for static images, and propose an architecture that can encode and predict Spatio-Temporal Affinity Fields (STAF) across a video sequence. In particular, we propose a novel temporal topology cross-linked across limbs which can consistently handle body motions of a wide range of magnitudes. Additionally, we make the overall approach recurrent in nature, where the network ingests STAF heatmaps from previous frames and estimates those for the current frame. Our approach uses only online inference and tracking, and is currently the fastest and the most accurate bottom-up approach that is runtime invariant to the number of people in the scene and accuracy invariant to input frame rate of camera. Running at $\sim$30 fps on a single GPU at single scale, it achieves highly competitive results on the PoseTrack benchmarks.
Configuring deep Spiking Neural Networks (SNNs) is an exciting research avenue for low power spike event based computation. However, the spike generation function is non-differentiable and therefore not directly compatible with the standard error backpropagation algorithm. In this paper, we introduce a new general backpropagation mechanism for learning synaptic weights and axonal delays which overcomes the problem of non-differentiability of the spike function and uses a temporal credit assignment policy for backpropagating error to preceding layers. We describe and release a GPU accelerated software implementation of our method which allows training both fully connected and convolutional neural network (CNN) architectures. Using our software, we compare our method against existing SNN based learning approaches and standard ANN to SNN conversion techniques and show that our method achieves state of the art performance for an SNN on the MNIST, NMNIST, DVS Gesture, and TIDIGITS datasets.
GPUs have limited memory and it is difficult to train wide and/or deep models that cause the training process to go out of memory. It is shown in this paper how an open source tool called Large Model Support (LMS) can utilize a high bandwidth NVLink connection between CPUs and GPUs to accomplish training of deep convolutional networks. LMS performs tensor swapping between CPU memory and GPU memory such that only a minimal number of tensors required in a training step are kept in the GPU memory. It is also shown how LMS can be combined with an MPI based distributed deep learning module to train models in a data-parallel fashion across multiple GPUs, such that each GPU is utilizing the CPU memory for tensor swapping. The hardware architecture that enables the high bandwidth GPU link with the CPU is discussed as well as the associated set of software tools that are available as the PowerAI package.
Fully-convolutional neural networks have achieved superior performance in a variety of image segmentation tasks. However, their training requires laborious manual annotation of large datasets, as well as acceleration by parallel processors with high-bandwidth memory, such as GPUs. We show that simple models can achieve competitive accuracy for organ segmentation on CT images when trained with extensive data augmentation, which leverages existing graphics hardware to quickly apply geometric and photometric transformations to 3D image data. On 3 mm^3 CT volumes, our GPU implementation is 2.6-8X faster than a widely-used CPU version, including communication overhead. We also show how to automatically generate training labels using rudimentary morphological operations, which are efficiently computed by 3D Fourier transforms. We combined fully-automatic labels for the lungs and bone with semi-automatic ones for the liver, kidneys and bladder, to create a dataset of 130 labeled CT scans. To achieve the best results from data augmentation, our model uses the intersection-over-union (IOU) loss function, a close relative of the Dice loss. We discuss its mathematical properties and explain why it outperforms the usual weighted cross-entropy loss for unbalanced segmentation tasks. We conclude that there is no unique IOU loss function, as the naive one belongs to a broad family of functions with the same essential properties. When combining data augmentation with the IOU loss, our model achieves a Dice score of 78-92% for each organ. The trained model, code and dataset will be made publicly available, to further medical imaging research.
Contextual information is vital in visual understanding problems, such as semantic segmentation and object detection. We propose a Criss-Cross Network (CCNet) for obtaining full-image contextual information in a very effective and efficient way. Concretely, for each pixel, a novel criss-cross attention module harvests the contextual information of all the pixels on its criss-cross path. By taking a further recurrent operation, each pixel can finally capture the full-image dependencies. Besides, a category consistent loss is proposed to enforce the criss-cross attention module to produce more discriminative features. Overall, CCNet is with the following merits: 1) GPU memory friendly. Compared with the non-local block, the proposed recurrent criss-cross attention module requires 11x less GPU memory usage. 2) High computational efficiency. The recurrent criss-cross attention significantly reduces FLOPs by about 85% of the non-local block. 3) The state-of-the-art performance. We conduct extensive experiments on semantic segmentation benchmarks including Cityscapes, ADE20K, human parsing benchmark LIP, instance segmentation benchmark COCO, video segmentation benchmark CamVid. In particular, our CCNet achieves the mIoU scores of 81.9%, 45.76% and 55.47% on the Cityscapes test set, the ADE20K validation set and the LIP validation set respectively, which are the new state-of-the-art results. The source codes are available at \url{https://github.com/speedinghzl/CCNet}.
One of the fundamental challenges in video object segmentation is to find an effective representation of the target and background appearance. The best performing approaches resort to extensive fine-tuning of a convolutional neural network for this purpose. Besides being prohibitively expensive, this strategy cannot be truly trained end-to-end since the online fine-tuning procedure is not integrated into the offline training of the network. To address these issues, we propose a network architecture that learns a powerful representation of the target and background appearance in a single forward pass. The introduced appearance module learns a probabilistic generative model of target and background feature distributions. Given a new image, it predicts the posterior class probabilities, providing a highly discriminative cue, which is processed in later network modules. Both the learning and prediction stages of our appearance module are fully differentiable, enabling true end-to-end training of the entire segmentation pipeline. Comprehensive experiments demonstrate the effectiveness of the proposed approach on three video object segmentation benchmarks. We close the gap to approaches based on online fine-tuning on DAVIS17, while operating at 15 FPS on a single GPU. Furthermore, our method outperforms all published approaches on the large-scale YouTube-VOS dataset.
There is a recent trend in handwritten text recognition with deep neural networks to replace 2D recurrent layers with 1D, and in some cases even completely remove the recurrent layers, relying on simple feed-forward convolutional only architectures. The most used type of recurrent layer is the Long-Short Term Memory (LSTM). The motivations to do so are many: there are few open-source implementations of 2D-LSTM, even fewer supporting GPU implementations (currently cuDNN only implements 1D-LSTM); 2D recurrences reduce the amount of computations that can be parallelized, and thus possibly increase the training/inference time; recurrences create global dependencies with respect to the input, and sometimes this may not be desirable. Many recent competitions were won by systems that employed networks that use 2D-LSTM layers. Most previous work that compared 1D or pure feed-forward architectures to 2D recurrent models have done so on simple datasets or did not fully optimize the "baseline" 2D model compared to the challenger model, which was dully optimized. In this work, we aim at a fair comparison between 2D and competing models and also extensively evaluate them on more complex datasets that are more representative of challenging "real-world" data, compared to "academic" datasets that are more restricted in their complexity. We aim at determining when and why the 1D and 2D recurrent models have different results. We also compare the results with a language model to assess if linguistic constraints do level the performance of the different networks. Our results show that for challenging datasets, 2D-LSTM networks still seem to provide the highest performances and we propose a visualization strategy to explain it.
We study two aspects of noisy computations during inference. The first aspect is how to mitigate their side effects for naturally trained deep learning systems. One of the motivations for looking into this problem is to reduce the high power cost of conventional computing of neural networks through the use of analog neuromorphic circuits. Traditional GPU/CPU-centered deep learning architectures exhibit bottlenecks in power-restricted applications (e.g., embedded systems). The use of specialized neuromorphic circuits, where analog signals passed through memory-cell arrays are sensed to accomplish matrix-vector multiplications, promises large power savings and speed gains but brings with it the problems of limited precision of computations and unavoidable analog noise. We manage to improve inference accuracy from 21.1% to 99.5% for MNIST images, from 29.9% to 89.1% for CIFAR10, and from 15.5% to 89.6% for MNIST stroke sequences with the presence of strong noise (with signal-to-noise power ratio being 0 dB) by noise-injected training and a voting method. This observation promises neural networks that are insensitive to inference noise, which reduces the quality requirements on neuromorphic circuits and is crucial for their practical usage. The second aspect is how to utilize the noisy inference as a defensive architecture against black-box adversarial attacks. During inference, by injecting proper noise to signals in the neural networks, the robustness of adversarially-trained neural networks against black-box attacks has been further enhanced by 0.5% and 1.13% for two adversarially trained models for MNIST and CIFAR10, respectively.
Deep neural networks have been widely used in numerous computer vision applications, particularly in face recognition. However, deploying deep neural network face recognition on mobile devices has recently become a trend but still limited since most high-accuracy deep models are both time and GPU consumption in the inference stage. Therefore, developing a lightweight deep neural network is one of the most practical solutions to deploy face recognition on mobile devices. Such the lightweight deep neural network requires efficient memory with small number of weights representation and low cost operators. In this paper, a novel deep neural network named MobiFace, a simple but effective approach, is proposed for productively deploying face recognition on mobile devices. The experimental results have shown that our lightweight MobiFace is able to achieve high performance with 99.73% on LFW database and 91.3% on large-scale challenging Megaface database. It is also eventually competitive against large-scale deep-networks face recognition while significant reducing computational time and memory consumption.