We present a theoretical and experimental investigation of the quantization problem for artificial neural networks. We provide a mathematical definition of quantized neural networks and analyze their approximation capabilities, showing in particular that any Lipschitz-continuous map defined on a hypercube can be uniformly approximated by a quantized neural network. We then focus on the regularization effect of additive noise on the arguments of multi-step functions inherent to the quantization of continuous variables. In particular, when the expectation operator is applied to a non-differentiable multi-step random function, and if the underlying probability density is differentiable (in either classical or weak sense), then a differentiable function is retrieved, with explicit bounds on its Lipschitz constant. Based on these results, we propose a novel gradient-based training algorithm for quantized neural networks that generalizes the straight-through estimator, acting on noise applied to the network's parameters. We evaluate our algorithm on the CIFAR-10 and ImageNet image classification benchmarks, showing state-of-the-art performance on AlexNet and MobileNetV2 for ternary networks.

In this paper, from a theoretical perspective, we study how powerful graph neural networks (GNNs) can be for learning approximation algorithms for combinatorial problems. To this end, we first establish a new class of GNNs that can solve strictly a wider variety of problems than existing GNNs. Then, we bridge the gap between GNN theory and the theory of distributed local algorithms to theoretically demonstrate that the most powerful GNN can learn approximation algorithms for the minimum dominating set problem and the minimum vertex cover problem with some approximation ratios and that no GNN can perform better than with these ratios. This paper is the first to elucidate approximation ratios of GNNs for combinatorial problems. Furthermore, we prove that adding coloring or weak-coloring to each node feature improves these approximation ratios. This indicates that preprocessing and feature engineering theoretically strengthen model capabilities.

Existing personalized dialogue models use human designed persona descriptions to improve dialogue consistency. Collecting such descriptions from existing dialogues is expensive and requires hand-crafted feature designs. In this paper, we propose to extend Model-Agnostic Meta-Learning (MAML)(Finn et al., 2017) to personalized dialogue learning without using any persona descriptions. Our model learns to quickly adapt to new personas by leveraging only a few dialogue samples collected from the same user, which is fundamentally different from conditioning the response on the persona descriptions. Empirical results on Persona-chat dataset (Zhang et al., 2018) indicate that our solution outperforms non-meta-learning baselines using automatic evaluation metrics, and in terms of human-evaluated fluency and consistency.

The paper studies how to release data about a critical infrastructure network (e.g., the power network or a transportation network) without disclosing sensitive information that can be exploited by malevolent agents, while preserving the realism of the network. It proposes a novel obfuscation mechanism that combines several privacy-preserving building blocks with a bi-level optimization model to significantly improve accuracy. The obfuscation is evaluated for both realism and privacy properties on real energy and transportation networks. Experimental results show the obfuscation mechanism substantially reduces the potential damage of an attack exploiting the released data to harm the real network.

Sliced Wasserstein metrics between probability measures solve the optimal transport (OT) problem on univariate projections, and average such maps across projections. The recent interest for the SW distance shows that much can be gained by looking at optimal maps between measures in smaller subspaces, as opposed to the curse-of-dimensionality price one has to pay in higher dimensions. Any transport estimated in a subspace remains, however, an object that can only be used in that subspace. We propose in this work two methods to extrapolate, from an transport map that is optimal on a subspace, one that is nearly optimal in the entire space. We prove that the best optimal transport plan that takes such "subspace detours" is a generalization of the Knothe-Rosenblatt transport. We show that these plans can be explicitly formulated when comparing Gaussians measures (between which the Wasserstein distance is usually referred to as the Bures or Fr\'echet distance). Building from there, we provide an algorithm to select optimal subspaces given pairs of Gaussian measures, and study scenarios in which that mediating subspace can be selected using prior information. We consider applications to NLP and evaluation of image quality (FID scores).

Convolutional neural networks (CNNs) have achieved remarkable performance in various fields, particularly in the domain of computer vision. However, why this architecture works well remains to be a mystery. In this work we move a small step toward understanding the success of CNNs by investigating the learning dynamics of a two-layer nonlinear convolutional neural network over some specific data distributions. Rather than the typical Gaussian assumption for input data distribution, we consider a more realistic setting that each data point (e.g. image) contains a specific pattern determining its class label. Within this setting, we both theoretically and empirically show that some convolutional filters will learn the key patterns in data and the norm of these filters will dominate during the training process with stochastic gradient descent. And with any high probability, when the number of iterations is sufficiently large, the CNN model could obtain 100% accuracy over the considered data distributions. Our experiments demonstrate that for practical image classification tasks our findings still hold to some extent.

Learning effective embedding has been proved to be useful in many real-world problems, such as recommender systems, search ranking and online advertisement. However, one of the challenges is data sparsity in learning large-scale item embedding, as users' historical behavior data are usually lacking or insufficient in an individual domain. In fact, user's behaviors from different domains regarding the same items are usually relevant. Therefore, we can learn complete user behaviors to alleviate the sparsity using complementary information from correlated domains. It is intuitive to model users' behaviors using graph, and graph neural networks (GNNs) have recently shown the great power for representation learning, which can be used to learn item embedding. However, it is challenging to transfer the information across domains and learn cross-domain representation using the existing GNNs. To address these challenges, in this paper, we propose a novel model - Deep Multi-Graph Embedding (DMGE) to learn cross-domain representation. Specifically, we first construct a multi-graph based on users' behaviors from different domains, and then propose a multi-graph neural network to learn cross-domain representation in an unsupervised manner. Particularly, we present a multiple-gradient descent optimizer for efficiently training the model. We evaluate our approach on various large-scale real-world datasets, and the experimental results show that DMGE outperforms other state-of-art embedding methods in various tasks.

Graph Neural Nets (GNNs) have received increasing attentions, partially due to their superior performance in many node and graph classification tasks. However, there is a lack of understanding on what they are learning and how sophisticated the learned graph functions are. In this work, we first propose Graph Feature Network (GFN), a simple lightweight neural net defined on a set of graph augmented features. We then propose a dissection of GNNs on graph classification into two parts: 1) the graph filtering, where graph-based neighbor aggregations are performed, and 2) the set function, where a set of hidden node features are composed for prediction. We prove that GFN can be derived by linearizing graph filtering part of GNNs, and leverage it to test the importance of the two parts separately. Empirically we perform evaluations on common graph classification benchmarks. To our surprise, we find that, despite the simplification, GFN could match or exceed the best accuracies produced by recently proposed GNNs, with a fraction of computation cost. Our results suggest that linear graph filtering with non-linear set function is powerful enough, and common graph classification benchmarks seem inadequate for testing advanced GNN variants.

Federated learning involves training statistical models in massive, heterogeneous networks. Naively minimizing an aggregate loss function in such a network may disproportionately advantage or disadvantage some of the devices. In this work, we propose q-Fair Federated Learning (q-FFL), a novel optimization objective inspired by resource allocation in wireless networks that encourages a more fair (i.e., lower-variance) accuracy distribution across devices in federated networks. To solve q-FFL, we devise a communication-efficient method, q-FedAvg, that is suited to federated networks. We validate both the effectiveness of q-FFL and the efficiency of q-FedAvg on a suite of federated datasets, and show that q-FFL (along with q-FedAvg) outperforms existing baselines in terms of the resulting fairness, flexibility, and efficiency.

The Euclidean scattering transform was introduced nearly a decade ago to improve the mathematical understanding of convolutional neural networks. Inspired by recent interest in geometric deep learning, which aims to generalize convolutional neural networks to manifold and graph-structured domains, we define a geometric scattering transform on manifolds. Similar to the Euclidean scattering transform, the geometric scattering transform is based on a cascade of wavelet filters and pointwise nonlinearities. It is invariant to local isometries and stable to certain types of diffeomorphisms. Empirical results demonstrate its utility on several geometric learning tasks. Our results generalize the deformation stability and local translation invariance of Euclidean scattering, and demonstrate the importance of linking the used filter structures to the underlying geometry of the data.