A major factor in the success of deep neural networks is the use of sophisticated architectures rather than the classical multilayer perceptron (MLP). Residual networks (ResNets) stand out among these powerful modern architectures. Previous works focused on the optimization advantages of deep ResNets over deep MLPs. In this paper, we show another distinction between the two models, namely, a tendency of ResNets to promote smoother interpolations than MLPs. We analyze this phenomenon via the neural tangent kernel (NTK) approach. First, we compute the NTK for a considered ResNet model and prove its stability during gradient descent training. Then, we show by various evaluation methodologies that the NTK of ResNet, and its kernel regression results, are smoother than the ones of MLP. The better smoothness observed in our analysis may explain the better generalization ability of ResNets and the practice of moderately attenuating the residual blocks.

Good training data is a prerequisite to develop useful ML applications. However, in many domains existing data sets cannot be shared due to privacy regulations (e.g., from medical studies). This work investigates a simple yet unconventional approach for anonymized data synthesis to enable third parties to benefit from such private data. We explore the feasibility of learning implicitly from unrealistic, task-relevant stimuli, which are synthesized by exciting the neurons of a trained deep neural network (DNN). As such, neuronal excitation serves as a pseudo-generative model. The stimuli data is used to train new classification models. Furthermore, we extend this framework to inhibit representations that are associated with specific individuals. We use sleep monitoring data from both an open and a large closed clinical study and evaluate whether (1) end-users can create and successfully use customized classification models for sleep apnea detection, and (2) the identity of participants in the study is protected. Extensive comparative empirical investigation shows that different algorithms trained on the stimuli are able generalize successfully on the same task as the original model. However, architectural and algorithmic similarity between new and original models play an important role in performance. For similar architectures, the performance is close to that of using the true data (e.g., Accuracy difference of 0.56\%, Kappa coefficient difference of 0.03-0.04). Further experiments show that the stimuli can to a large extent successfully anonymize participants of the clinical studies.

As neural networks make their way into safety-critical systems, where misbehavior can lead to catastrophes, there is a growing interest in certifying the equivalence of two structurally similar neural networks. For example, compression techniques are often used in practice for deploying trained neural networks on computationally- and energy-constrained devices, which raises the question of how faithfully the compressed network mimics the original network. Unfortunately, existing methods either focus on verifying a single network or rely on loose approximations to prove the equivalence of two networks. Due to overly conservative approximation, differential verification lacks scalability in terms of both accuracy and computational cost. To overcome these problems, we propose NeuroDiff, a symbolic and fine-grained approximation technique that drastically increases the accuracy of differential verification while achieving many orders-of-magnitude speedup. NeuroDiff has two key contributions. The first one is new convex approximations that more accurately bound the difference neurons of two networks under all possible inputs. The second one is judicious use of symbolic variables to represent neurons whose difference bounds have accumulated significant error. We also find that these two techniques are complementary, i.e., when combined, the benefit is greater than the sum of their individual benefits. We have evaluated NeuroDiff on a variety of differential verification tasks. Our results show that NeuroDiff is up to 1000X faster and 5X more accurate than the state-of-the-art tool.

Deep learning models are shown to be vulnerable to adversarial examples. Though adversarial training can enhance model robustness, typical approaches are computationally expensive. Recent works proposed to transfer the robustness to adversarial attacks across different tasks or models with soft labels.Compared to soft labels, feature contains rich semantic information and holds the potential to be applied to different downstream tasks. In this paper, we propose a novel approach called Guided Adversarial Contrastive Distillation (GACD), to effectively transfer adversarial robustness from teacher to student with features. We first formulate this objective as contrastive learning and connect it with mutual information. With a well-trained teacher model as an anchor, students are expected to extract features similar to the teacher. Then considering the potential errors made by teachers, we propose sample reweighted estimation to eliminate the negative effects from teachers. With GACD, the student not only learns to extract robust features, but also captures structural knowledge from the teacher. By extensive experiments evaluating over popular datasets such as CIFAR-10, CIFAR-100 and STL-10, we demonstrate that our approach can effectively transfer robustness across different models and even different tasks, and achieve comparable or better results than existing methods. Besides, we provide a detailed analysis of various methods, showing that students produced by our approach capture more structural knowledge from teachers and learn more robust features under adversarial attacks.

Graph property prediction is drawing increasing attention in the recent years due to the fact that graphs are one of the most general data structures since they can contain an arbitrary number of nodes and connections between them, and it is the backbone for many different tasks like classification and regression on such kind of data (networks, molecules, knowledge bases, ...). We introduce a novel generalized global pooling layer to mitigate the information loss that typically occurs at the Readout phase in Message-Passing Neural Networks. This novel layer is parametrized by two values ($\beta$ and $p$) which can optionally be learned, and the transformation it performs can revert to several already popular readout functions (mean, max and sum) under certain settings, which can be specified. To showcase the superior expressiveness and performance of this novel technique, we test it in a popular graph property prediction task by taking the current best-performing architecture and using our readout layer as a drop-in replacement and we report new state of the art results. The code to reproduce the experiments can be accessed here: https://github.com/EricAlcaide/generalized-readout-phase

Leverage score sampling is a powerful technique that originates from theoretical computer science, which can be used to speed up a large number of fundamental questions, e.g. linear regression, linear programming, semi-definite programming, cutting plane method, graph sparsification, maximum matching and max-flow. Recently, it has been shown that leverage score sampling helps to accelerate kernel methods [Avron, Kapralov, Musco, Musco, Velingker and Zandieh 17]. In this work, we generalize the results in [Avron, Kapralov, Musco, Musco, Velingker and Zandieh 17] to a broader class of kernels. We further bring the leverage score sampling into the field of deep learning theory. $\bullet$ We show the connection between the initialization for neural network training and approximating the neural tangent kernel with random features. $\bullet$ We prove the equivalence between regularized neural network and neural tangent kernel ridge regression under the initialization of both classical random Gaussian and leverage score sampling.

Purpose: The goal of this research was to develop a deep network architecture that achieves fully-automated radiologist-level segmentation of breast tumors in MRI. Materials and Methods: We leveraged 38,229 clinical MRI breast exams collected retrospectively from women aged 12-94 (mean age 54) who presented between 2002 and 2014 at a single clinical site. The training set for the network consisted of 2,555 malignant breasts that were segmented in 2D by experienced radiologists, as well as 60,108 benign breasts that served as negative controls. The test set consisted of 250 exams with tumors segmented independently by four radiologists. We selected among several 3D deep convolutional neural network architectures, input modalities and harmonization methods. The outcome measure was the Dice score for 2D segmentation, and was compared between the network and radiologists using the Wilcoxon signed-rank test and the TOST procedure. Results: The best-performing network on the training set was a volumetric U-Net with contrast enhancement dynamic as input and with intensity normalized for each exam. In the test set the median Dice score of this network was 0.77. The performance of the network was equivalent to that of the radiologists (TOST procedure with radiologist performance of 0.69-0.84 as equivalence bounds: p = 5e-10 and p = 2e-5, respectively; N = 250) and compares favorably with published state of the art (0.6-0.77). Conclusion: When trained on a dataset of over 60 thousand breasts, a volumetric U-Net performs as well as expert radiologists at segmenting malignant breast lesions in MRI.

Knowledge of the hidden factors that determine particular system dynamics is crucial for both explaining them and pursuing goal-directed, interventional actions. The inference of these factors without supervision given time series data remains an open challenge. Here, we focus on spatio-temporal processes, including wave propagations and weather dynamics, and assume that universal causes (e.g. physics) apply throughout space and time. We apply a novel DIstributed, Spatio-Temporal graph Artificial Neural network Architecture, DISTANA, which learns a generative model in such domains. DISTANA requires fewer parameters, and yields more accurate predictions than temporal convolutional neural networks and other related approaches on a 2D circular wave prediction task. We show that DISTANA, when combined with a retrospective latent state inference principle called active tuning, can reliably derive hidden local causal factors. In a current weather prediction benchmark, DISTANA infers our planet's land-sea mask solely by observing temperature dynamics and uses the self inferred information to improve its own prediction of temperature. We are convinced that the retrospective inference of latent states in generative RNN architectures will play an essential role in future research on causal inference and explainable systems.

In this work, we propose DiffWave, a versatile diffusion probabilistic model for conditional and unconditional waveform generation. The model is non-autoregressive, and converts the white noise signal into structured waveform through a Markov chain with a constant number of steps at synthesis. It is efficiently trained by optimizing a variant of variational bound on the data likelihood. DiffWave produces high-fidelity audio in different waveform generation tasks, including neural vocoding conditioned on mel spectrogram, class-conditional generation, and unconditional generation. We demonstrate that DiffWave matches a strong WaveNet vocoder in terms of speech quality (MOS: 4.44 versus 4.43), while synthesizing orders of magnitude faster. In particular, it significantly outperforms autoregressive and GAN-based waveform models in the challenging unconditional generation task in terms of audio quality and sample diversity from various automatic and human evaluations. Deep generative models have produced high-fidelity raw audio in speech synthesis and music generation. In previous work, likelihood-based models, including autoregressive models (van den Oord et al., 2016; Kalchbrenner et al., 2018; Mehri et al., 2017) and flow-based models (Prenger et al., 2019; Ping et al., 2020; Kim et al., 2019), have predominated in audio synthesis because of the simple training objective and superior ability of modeling the fine details of waveform in real data.

In order to further improve the search efficiency of Neural Architecture Search (NAS), we propose B-DARTS, a novel pipeline combining broad scalable architecture with Confident Learning Rate (CLR). In B-DARTS, Broad Convolutional Neural Network (BCNN) is employed as the scalable architecture for DARTS, a popular differentiable NAS approach. On one hand, BCNN is a broad scalable architecture whose topology achieves two advantages compared with the deep one, mainly including faster single-step training speed and higher memory efficiency (i.e. larger batch size for architecture search), which are all contributed to the search efficiency improvement of NAS. On the other hand, DARTS discovers the optimal architecture by gradient-based optimization algorithm, which benefits from two superiorities of BCNN simultaneously. Similar to vanilla DARTS, B-DARTS also suffers from the performance collapse issue, where those weight-free operations are prone to be selected by the search strategy. Therefore, we propose CLR, that considers the confidence of gradient for architecture weights update increasing with the training time of over-parameterized model, to mitigate the above issue. Experimental results on CIFAR-10 and ImageNet show that 1) B-DARTS delivers state-of-the-art efficiency of 0.09 GPU day using first order approximation on CIFAR-10; 2) the learned architecture by B-DARTS achieves competitive performance using state-of-the-art composite multiply-accumulate operations and parameters on ImageNet; and 3) the proposed CLR is effective for performance collapse issue alleviation of both B-DARTS and DARTS.