Adversarial training is a common approach to improving the robustness of deep neural networks against adversarial examples. In this work, we propose a novel regularization approach as an alternative. To derive the regularizer, we formulate the adversarial robustness problem under the robust optimization framework and approximate the loss function using a second-order Taylor series expansion. Our proposed second-order adversarial regularizer (SOAR) is an upper bound based on the Taylor approximation of the inner-max in the robust optimization objective. We empirically show that the proposed method improves the robustness of networks on the CIFAR-10 dataset.

We formalize an equivalence between two popular methods for Bayesian inference: Stein variational gradient descent (SVGD) and black-box variational inference (BBVI). In particular, we show that BBVI corresponds precisely to SVGD when the kernel is the neural tangent kernel. Furthermore, we interpret SVGD and BBVI as kernel gradient flows; we do this by leveraging the recent perspective that views SVGD as a gradient flow in the space of probability distributions and showing that BBVI naturally motivates a Riemannian structure on that space. We observe that kernel gradient flow also describes dynamics found in the training of generative adversarial networks (GANs). This work thereby unifies several existing techniques in variational inference and generative modeling and identifies the kernel as a fundamental object governing the behavior of these algorithms, motivating deeper analysis of its properties.

The objective of this study is to predict suicidal and non-suicidal deaths from DNA methylation data using a modern machine learning algorithm. We used support vector machines to classify existing secondary data consisting of normalized values of methylated DNA probe intensities from tissues of two cortical brain regions to distinguish suicide cases from control cases. Before classification, we employed Principal component analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE) to reduce the dimension of the data. In comparison to PCA, the modern data visualization method t-SNE performs better in dimensionality reduction. t-SNE accounts for the possible non-linear patterns in low-dimensional data. We applied four-fold cross-validation in which the resulting output from t-SNE was used as training data for the Support Vector Machine (SVM). Despite the use of cross-validation, the nominally perfect prediction of suicidal deaths for BA11 data suggests possible over-fitting of the model. The study also may have suffered from 'spectrum bias' since the individuals were only studied from two extreme scenarios. This research constitutes a baseline study for classifying suicidal and non-suicidal deaths from DNA methylation data. Future studies with larger sample size, while possibly incorporating methylation data from living individuals, may reduce the bias and improve the accuracy of the results.

This study uses stacked generalization, which is a two-step process of combining machine learning methods, called meta or super learners, for improving the performance of algorithms in step one (by minimizing the error rate of each individual algorithm to reduce its bias in the learning set) and then in step two inputting the results into the meta learner with its stacked blended output (demonstrating improved performance with the weakest algorithms learning better). The method is essentially an enhanced cross-validation strategy. Although the process uses great computational resources, the resulting performance metrics on resampled fraud data show that increased system cost can be justified. A fundamental key to fraud data is that it is inherently not systematic and, as of yet, the optimal resampling methodology has not been identified. Building a test harness that accounts for all permutations of algorithm sample set pairs demonstrates that the complex, intrinsic data structures are all thoroughly tested. Using a comparative analysis on fraud data that applies stacked generalizations provides useful insight needed to find the optimal mathematical formula to be used for imbalanced fraud data sets.

As machine learning (ML) has seen increasing adoption in safety-critical domains (e.g., autonomous vehicles), the reliability of ML systems has also grown in importance. While prior studies have proposed techniques to enable efficient error-resilience techniques (e.g., selective instruction duplication), a fundamental requirement for realizing these techniques is a detailed understanding of the application's resilience. In this work, we present TensorFI, a high-level fault injection (FI) framework for TensorFlow-based applications. TensorFI is able to inject both hardware and software faults in general TensorFlow programs. TensorFI is a configurable FI tool that is flexible, easy to use, and portable. It can be integrated into existing TensorFlow programs to assess their resilience for different fault types (e.g., faults in particular operators). We use TensorFI to evaluate the resilience of 12 ML programs, including DNNs used in the autonomous vehicle domain. Our tool is publicly available at https://github.com/DependableSystemsLab/TensorFI.

The bulk of universal algorithms in the online convex optimisation literature are variants of the Hedge (exponential weights) algorithm on the simplex. While these algorithms extend to polytope domains by assigning weights to the vertices, this process is computationally unfeasible for many important classes of polytopes where the number $V$ of vertices depends exponentially on the dimension $d$. In this paper we show the Subgradient algorithm is universal, meaning it has $O(\sqrt N)$ regret in the antagonistic setting and $O(1)$ pseudo-regret in the i.i.d setting, with two main advantages over Hedge: (1) The update step is more efficient as the action vectors have length only $d$ rather than $V$; and (2) Subgradient gives better performance if the cost vectors satisfy Euclidean rather than sup-norm bounds. This paper extends the authors' recent results for Subgradient on the simplex. We also prove the same $O(\sqrt N)$ and $O(1)$ bounds when the domain is the unit ball. To the authors' knowledge this is the first instance of these bounds on a domain other than a polytope.

Domain adaptation aims to exploit a label-rich source domain for learning classifiers in a different label-scarce target domain. It is particularly challenging when there are significant divergences between the two domains. In the paper, we propose a novel unsupervised domain adaptation method based on progressive domain augmentation. The proposed method generates virtual intermediate domains via domain interpolation, progressively augments the source domain and bridges the source-target domain divergence by conducting multiple subspace alignment on the Grassmann manifold. We conduct experiments on multiple domain adaptation tasks and the results shows the proposed method achieves the state-of-the-art performance.

Social media has greatly enabled people to participate in online activities at an unprecedented rate. However, this unrestricted access also exacerbates the spread of misinformation and fake news online which might cause confusion and chaos unless being detected early for its mitigation. Given the rapidly evolving nature of news events and the limited amount of annotated data, state-of-the-art systems on fake news detection face challenges due to the lack of large numbers of annotated training instances that are hard to come by for early detection. In this work, we exploit multiple weak signals from different sources given by user and content engagements (referred to as weak social supervision), and their complementary utilities to detect fake news. We jointly leverage the limited amount of clean data along with weak signals from social engagements to train deep neural networks in a meta-learning framework to estimate the quality of different weak instances. Experiments on realworld datasets demonstrate that the proposed framework outperforms state-of-the-art baselines for early detection of fake news without using any user engagements at prediction time.

Non-autoregressive machine translation models significantly speed up decoding by allowing for parallel prediction of the entire target sequence. However, modeling word order is more challenging due to the lack of autoregressive factors in the model. This difficultly is compounded during training with cross entropy loss, which can highly penalize small shifts in word order. In this paper, we propose aligned cross entropy (AXE) as an alternative loss function for training of non-autoregressive models. AXE uses a differentiable dynamic program to assign loss based on the best possible monotonic alignment between target tokens and model predictions. AXE-based training of conditional masked language models (CMLMs) substantially improves performance on major WMT benchmarks, while setting a new state of the art for non-autoregressive models.

We propose orthogonal inductive matrix completion (OMIC), an interpretable model composed of a sum of matrix completion terms, each with orthonormal side information. We can inject prior knowledge about the eigenvectors of the ground truth matrix, whilst maintaining the representation capability of the model. We present a provably converging algorithm that optimizes all components of the model simultaneously, using nuclear-norm regularisation. Our method is backed up by \textit{distribution-free} learning guarantees that improve with the quality of the injected knowledge. As a special case of our general framework, we study a model consisting of a sum of user and item biases (generic behaviour), a non-inductive term (specific behaviour), and an inductive term using side information. Our theoretical analysis shows that $\epsilon$-recovering the ground truth matrix requires at most $O\left( \frac{n+m+(\sqrt{n}+\sqrt{m})mn \sqrt{r}C}{\epsilon^2}\right)$ entries, where $r$ is the rank of the ground truth matrix. We analyse the performance of OMIC on several synthetic and real datasets. On synthetic datasets with a sliding scale of user bias relevance, we show that OMIC better adapts to different regimes than other methods and can recover the ground truth. On real life datasets containing user/items recommendations and relevant side information, we find that OMIC surpasses the state of the art, with the added benefit of greater interpretability.