Transformers are ubiquitous in Natural Language Processing (NLP) tasks, but they are difficult to be deployed on hardware due to the intensive computation. To enable low-latency inference on resource-constrained hardware platforms, we propose to design Hardware-Aware Transformers (HAT) with neural architecture search. We first construct a large design space with $\textit{arbitrary encoder-decoder attention}$ and $\textit{heterogeneous layers}$. Then we train a $\textit{SuperTransformer}$ that covers all candidates in the design space, and efficiently produces many $\textit{SubTransformers}$ with weight sharing. Finally, we perform an evolutionary search with a hardware latency constraint to find a specialized $\textit{SubTransformer}$ dedicated to run fast on the target hardware. Extensive experiments on four machine translation tasks demonstrate that HAT can discover efficient models for different hardware (CPU, GPU, IoT device). When running WMT'14 translation task on Raspberry Pi-4, HAT can achieve $\textbf{3}\times$ speedup, $\textbf{3.7}\times$ smaller size over baseline Transformer; $\textbf{2.7}\times$ speedup, $\textbf{3.6}\times$ smaller size over Evolved Transformer with $\textbf{12,041}\times$ less search cost and no performance loss. HAT code is https://github.com/mit-han-lab/hardware-aware-transformers.git

Computer-aided diagnosis (CAD) is becoming a prominent approach to assist clinicians spanning across multiple fields. These automated systems take advantage of various computer vision (CV) procedures, as well as artificial intelligence (AI) techniques, so that a diagnosis of a given image (e.g., computed tomography and ultrasound) can be formulated. Advances in both areas (CV and AI) are enabling ever increasing performances of CAD systems, which can ultimately avoid performing invasive procedures such as fine-needle aspiration. In this study, we focus on thyroid ultrasonography to present a novel knowledge-driven classification framework. The proposed system leverages cues provided by an ensemble of experts, in order to guide the learning phase of a densely connected convolutional network (DenseNet). The ensemble is composed by various networks pretrained on ImageNet, including AlexNet, ResNet, VGG, and others, so that previously computed feature parameters could be used to create ultrasonography domain experts via transfer learning, decreasing, moreover, the number of samples required for training. To validate the proposed method, extensive experiments were performed, providing detailed performances for both the experts ensemble and the knowledge-driven DenseNet. The obtained results, show how the the proposed system can become a great asset when formulating a diagnosis, by leveraging previous knowledge derived from a consult.

Physics-Informed Neural Networks (PINNs) can be regarded as general-purpose PDE solvers, but it might be slow to train PINNs on particular problems, and there is no theoretical guarantee of corresponding error bounds. In this manuscript, we propose a variant called Prior Dictionary based Physics-Informed Neural Networks (PD-PINNs). Equipped with task-dependent dictionaries, PD-PINNs enjoy enhanced representation power on the tasks, which helps to capture features provided by dictionaries so that the proposed neural networks can achieve faster convergence in the process of training. In various numerical simulations, compared with existing PINN methods, combining prior dictionaries can significantly enhance convergence speed. In terms of theory, we obtain the error bounds applicable to PINNs and PD-PINNs for solving elliptic partial differential equations of second order. It is proved that under certain mild conditions, the prediction error made by neural networks can be bounded by expected loss of PDEs and boundary conditions.

Bokeh effect is used in photography to capture images where the closer objects look sharp and every-thing else stays out-of-focus. Bokeh photos are generally captured using Single Lens Reflex cameras using shallow depth-of-field. Most of the modern smartphones can take bokeh images by leveraging dual rear cameras or a good auto-focus hardware. However, for smartphones with single-rear camera without a good auto-focus hardware, we have to rely on software to generate bokeh images. This kind of system is also useful to generate bokeh effect in already captured images. In this paper, an end-to-end deep learning framework is proposed to generate high-quality bokeh effect from images. The original image and different versions of smoothed images are blended to generate Bokeh effect with the help of a monocular depth estimation network. The proposed approach is compared against a saliency detection based baseline and a number of approaches proposed in AIM 2019 Challenge on Bokeh Effect Synthesis. Extensive experiments are shown in order to understand different parts of the proposed algorithm. The network is lightweight and can process an HD image in 0.03 seconds. This approach ranked second in AIM 2019 Bokeh effect challenge-Perceptual Track.

Having been collected for its primary purpose in patient care, Observational Health Data (OHD) can further benefit patient well-being by sustaining the development of health informatics. However, the potential for secondary usage of OHD continues to be hampered by the fiercely private nature of patient-related data. Generative Adversarial Networks (GAN) have Generative Adversarial Networks (GAN) have recently emerged as a groundbreaking approach to efficiently learn generative models that produce realistic Synthetic Data (SD). However, the application of GAN to OHD seems to have been lagging in comparison to other fields. We conducted a review of GAN algorithms for OHD in the published literature, and report our findings here.

Bayesian neural Networks (BNNs) are a promising method of obtaining statistical uncertainties for neural network predictions but with a higher computational overhead which can limit their practical usage. This work explores the use of high performance computing with distributed training to address the challenges of training BNNs at scale. We present a performance and scalability comparison of training the VGG-16 and Resnet-18 models on a Cray-XC40 cluster. We demonstrate that network pruning can speed up inference without accuracy loss and provide an open source software package, {\it{BPrune}} to automate this pruning. For certain models we find that pruning up to 80\% of the network results in only a 7.0\% loss in accuracy. With the development of new hardware accelerators for Deep Learning, BNNs are of considerable interest for benchmarking performance. This analysis of training a BNN at scale outlines the limitations and benefits compared to a conventional neural network.

Recognizing sounds is a key aspect of computational audio scene analysis and machine perception. In this paper, we advocate that sound recognition is inherently a multi-modal audiovisual task in that it is easier to differentiate sounds using both the audio and visual modalities as opposed to one or the other. We present an audiovisual fusion model that learns to recognize sounds from weakly labeled video recordings. The proposed fusion model utilizes an attention mechanism to dynamically combine the outputs of the individual audio and visual models. Experiments on the large scale sound events dataset, AudioSet, demonstrate the efficacy of the proposed model, which outperforms the single-modal models, and state-of-the-art fusion and multi-modal models. We achieve a mean Average Precision (mAP) of 46.16 on Audioset, outperforming prior state of the art by approximately +4.35 mAP (relative: 10.4%).

In machine learning, the performance of a classifier depends on both the classifier model and the dataset. For a specific neural network classifier, the training process varies with the training set used; some training data make training accuracy fast converged to high values, while some data may lead to slowly converged to lower accuracy. To quantify this phenomenon, we created the Distance-based Separability Index (DSI), which is independent of the classifier model, to measure the separability of datasets. In this paper, we consider the situation where different classes of data are mixed together in the same distribution is most difficult for classifiers to separate, and we show that the DSI can indicate whether data belonging to different classes have similar distributions. When comparing our proposed approach with several existing separability/complexity measures using synthetic and real datasets, the results show the DSI is an effective separability measure. We also discussed possible applications of the DSI in the fields of data science, machine learning, and deep learning.

Machine learning (ML), especially deep neural networks (DNNs) have been widely used in various applications, including several safety-critical ones (e.g. autonomous driving). As a result, recent research about adversarial examples has raised great concerns. Such adversarial attacks can be achieved by adding a small magnitude of perturbation to the input to mislead model prediction. While several whitebox attacks have demonstrated their effectiveness, which assume that the attackers have full access to the machine learning models; blackbox attacks are more realistic in practice. In this paper, we propose a Query-Efficient Boundary-based blackbox Attack (QEBA) based only on model's final prediction labels. We theoretically show why previous boundary-based attack with gradient estimation on the whole gradient space is not efficient in terms of query numbers, and provide optimality analysis for our dimension reduction-based gradient estimation. On the other hand, we conducted extensive experiments on ImageNet and CelebA datasets to evaluate QEBA. We show that compared with the state-of-the-art blackbox attacks, QEBA is able to use a smaller number of queries to achieve a lower magnitude of perturbation with 100% attack success rate. We also show case studies of attacks on real-world APIs including MEGVII Face++ and Microsoft Azure.

In view of the undisputed success of neural networks and due to the remarkable recent improvements in their ability to solve a huge variety of practical problems, the development of a satisfying and rigorous mathematical understanding of their performance is one of the main challenges in the field of learning theory. Against this background, we study the expressive power of neural networks through the example of the classical NP-hard Knapsack Problem. Our main contribution is a class of recurrent neural networks (RNNs) with rectified linear units that are iteratively applied to each item of a Knapsack instance and thereby compute optimal or provably good solution values. In order to find optimum Knapsack solutions, an RNN of depth four and width depending quadratically on the profit of an optimum Knapsack solution is sufficient. We also prove the following tradeoff between the size of an RNN and the quality of the computed Knapsack solution: For Knapsack instances consisting of $n$ items, an RNN of depth five and width $w$ computes a solution of value at least $1-\mathcal{O}(n^2/\sqrt{w})$ times the optimum solution value. Our results build upon a dynamic programming formulation of the Knapsack Problem as well as a careful rounding of profit values that is also at the core of the well-known fully polynomial-time approximation scheme for the Knapsack Problem. Finally, we point out that similar results can be achieved for other optimization problems that can be solved by dynamic programming, such as, e.g., various Shortest Path Problems and the Longest Common Subsequence Problem.