This paper analyzes the trajectories of stochastic gradient descent (SGD) to help understand the algorithm's convergence properties in non-convex problems. We first show that the sequence of iterates generated by SGD remains bounded and converges with probability $1$ under a very broad range of step-size schedules. Subsequently, going beyond existing positive probability guarantees, we show that SGD avoids strict saddle points/manifolds with probability $1$ for the entire spectrum of step-size policies considered. Finally, we prove that the algorithm's rate of convergence to Hurwicz minimizers is $\mathcal{O}(1/n^{p})$ if the method is employed with a $\Theta(1/n^p)$ step-size schedule. This provides an important guideline for tuning the algorithm's step-size as it suggests that a cool-down phase with a vanishing step-size could lead to faster convergence; we demonstrate this heuristic using ResNet architectures on CIFAR.

Exploratory landscape analysis (ELA) supports supervised learning approaches for automated algorithm selection and configuration by providing sets of features that quantify the most relevant characteristics of the optimization problem at hand. In black-box optimization, where an explicit problem representation is not available, the feature values need to be approximated from a small number of sample points. In practice, uniformly sampled random point sets and Latin hypercube constructions are commonly used sampling strategies. In this work, we analyze how the sampling method and the sample size influence the quality of the feature value approximations and how this quality impacts the accuracy of a standard classification task. While, not unexpectedly, increasing the number of sample points gives more robust estimates for the feature values, to our surprise we find that the feature value approximations for different sampling strategies do not converge to the same value. This implies that approximated feature values cannot be interpreted independently of the underlying sampling strategy. As our classification experiments show, this also implies that the feature approximations used for training a classifier must stem from the same sampling strategy as those used for the actual classification tasks. As a side result we show that classifiers trained with feature values approximated by Sobol' sequences achieve higher accuracy than any of the standard sampling techniques. This may indicate improvement potential for ELA-trained machine learning models.

Graph neural networks have attracted wide attentions to enable representation learning of graph data in recent works. In complement to graph convolution operators, graph pooling is crucial for extracting hierarchical representation of graph data. However, most recent graph pooling methods still fail to efficiently exploit the geometry of graph data. In this paper, we propose a novel graph pooling strategy that leverages node proximity to improve the hierarchical representation learning of graph data with their multi-hop topology. Node proximity is obtained by harmonizing the kernel representation of topology information and node features. Implicit structure-aware kernel representation of topology information allows efficient graph pooling without explicit eigendecomposition of the graph Laplacian. Similarities of node signals are adaptively evaluated with the combination of the affine transformation and kernel trick using the Gaussian RBF function. Experimental results demonstrate that the proposed graph pooling strategy is able to achieve state-of-the-art performance on a collection of public graph classification benchmark datasets.

Conditional correlation networks, within Gaussian Graphical Models (GGM), are widely used to describe the direct interactions between the components of a random vector. In the case of an unlabelled Heterogeneous population, Expectation Maximisation (EM) algorithms for Mixtures of GGM have been proposed to estimate both each sub-population's graph and the class labels. However, we argue that, with most real data, class affiliation cannot be described with a Mixture of Gaussian, which mostly groups data points according to their geometrical proximity. In particular, there often exists external co-features whose values affect the features' average value, scattering across the feature space data points belonging to the same sub-population. Additionally, if the co-features' effect on the features is Heterogeneous, then the estimation of this effect cannot be separated from the sub-population identification. In this article, we propose a Mixture of Conditional GGM (CGGM) that subtracts the heterogeneous effects of the co-features to regroup the data points into sub-population corresponding clusters. We develop a penalised EM algorithm to estimate graph-sparse model parameters. We demonstrate on synthetic and real data how this method fulfils its goal and succeeds in identifying the sub-populations where the Mixtures of GGM are disrupted by the effect of the co-features.

Recurrent neural networks (RNNs) are brain-inspired models widely used in machine learning for analyzing sequential data. The present work is a contribution towards a deeper understanding of how RNNs process input signals using the response theory from nonequilibrium statistical mechanics. For a class of continuous-time stochastic RNNs (SRNNs) driven by an input signal, we derive a Volterra type series representation for their output. This representation is interpretable and disentangles the input signal from the SRNN architecture. The kernels of the series are certain recursively defined correlation functions with respect to the unperturbed dynamics that completely determine the output. Exploiting connections of this representation and its implications to rough paths theory, we identify a universal feature -- the response feature, which turns out to be the signature of tensor product of the input signal and a natural support basis. In particular, we show that the SRNNs can be viewed as kernel machines operating on a reproducing kernel Hilbert space associated with the response feature.

The gradient boosting machine is a powerful ensemble-based machine learning method for solving regression problems. However, one of the difficulties of its using is a possible discontinuity of the regression function, which arises when regions of training data are not densely covered by training points. In order to overcome this difficulty and to reduce the computational complexity of the gradient boosting machine, we propose to apply the partially randomized trees which can be regarded as a special case of the extremely randomized trees applied to the gradient boosting. The gradient boosting machine with the partially randomized trees is illustrated by means of many numerical examples using synthetic and real data.

Self-training is a classical approach in semi-supervised learning which is successfully applied to a variety of machine learning problems. Self-training algorithm generates pseudo-labels for the unlabeled examples and progressively refines these pseudo-labels which hopefully coincides with the actual labels. This work provides theoretical insights into self-training algorithm with a focus on linear classifiers. We first investigate Gaussian mixture models and provide a sharp non-asymptotic finite-sample characterization of the self-training iterations. Our analysis reveals the provable benefits of rejecting samples with low confidence and demonstrates that self-training iterations gracefully improve the model accuracy even if they do get stuck in sub-optimal fixed points. We then demonstrate that regularization and class margin (i.e. separation) is provably important for the success and lack of regularization may prevent self-training from identifying the core features in the data. Finally, we discuss statistical aspects of empirical risk minimization with self-training for general distributions. We show how a purely unsupervised notion of generalization based on self-training based clustering can be formalized based on cluster margin. We then establish a connection between self-training based semi-supervision and the more general problem of learning with heterogenous data and weak supervision.

Recent work has highlighted the vulnerability of many deep machine learning models to adversarial examples. It attracts increasing attention to adversarial attacks, which can be used to evaluate the security and robustness of models before they are deployed. However, to our best knowledge, there is no specific research on the adversarial attacks for multi-view deep models. This paper proposes two multi-view attack strategies, two-stage attack (TSA) and end-to-end attack (ETEA). With the mild assumption that the single-view model on which the target multi-view model is based is known, we first propose the TSA strategy. The main idea of TSA is to attack the multi-view model with adversarial examples generated by attacking the associated single-view model, by which state-of-the-art single-view attack methods are directly extended to the multi-view scenario. Then we further propose the ETEA strategy when the multi-view model is provided publicly. The ETEA is applied to accomplish direct attacks on the target multi-view model, where we develop three effective multi-view attack methods. Finally, based on the fact that adversarial examples generalize well among different models, this paper takes the adversarial attack on the multi-view convolutional neural network as an example to validate that the effectiveness of the proposed multi-view attacks. Extensive experimental results demonstrate that our multi-view attack strategies are capable of attacking the multi-view deep models, and we additionally find that multi-view models are more robust than single-view models.

Recommendation is a prevalent application of machine learning that affects many users; therefore, it is important for recommender models to be accurate and interpretable. In this work, we propose a method to both interpret and augment the predictions of black-box recommender systems. In particular, we propose to interpret feature interactions from a source recommender model and explicitly encode these interactions in a target recommender model, where both source and target models are black-boxes. By not assuming the structure of the recommender system, our approach can be used in general settings. In our experiments, we focus on a prominent use of machine learning recommendation: ad-click prediction. We found that our interaction interpretations are both informative and predictive, e.g., significantly outperforming existing recommender models. What's more, the same approach to interpret interactions can provide new insights into domains even beyond recommendation, such as text and image classification.

Effective optimization is essential for real-world interactive systems to provide a satisfactory user experience in response to changing user behavior. However, it is often challenging to find an objective to optimize for interactive systems (e.g., policy learning in task-oriented dialog systems). Generally, such objectives are manually crafted and rarely capture complex user needs in an accurate manner. We propose an approach that infers the objective directly from observed user interactions. These inferences can be made regardless of prior knowledge and across different types of user behavior. We introduce Interactive System Optimizer (ISO), a novel algorithm that uses these inferred objectives for optimization. Our main contribution is a new general principled approach to optimizing interactive systems using data-driven objectives. We demonstrate the high effectiveness of ISO over several simulations.