A number of studies have found that today's Visual Question Answering (VQA) models are heavily driven by superficial correlations in the training data and lack sufficient image grounding. To encourage development of models geared towards the latter, we propose a new setting for VQA where for every question type, train and test sets have different prior distributions of answers. Specifically, we present new splits of the VQA v1 and VQA v2 datasets, which we call Visual Question Answering under Changing Priors (VQA-CP v1 and VQA-CP v2 respectively). First, we evaluate several existing VQA models under this new setting and show that their performance degrades significantly compared to the original VQA setting. Second, we propose a novel Grounded Visual Question Answering model (GVQA) that contains inductive biases and restrictions in the architecture specifically designed to prevent the model from 'cheating' by primarily relying on priors in the training data. Specifically, GVQA explicitly disentangles the recognition of visual concepts present in the image from the identification of plausible answer space for a given question, enabling the model to more robustly generalize across different distributions of answers. GVQA is built off an existing VQA model -- Stacked Attention Networks (SAN). Our experiments demonstrate that GVQA significantly outperforms SAN on both VQA-CP v1 and VQA-CP v2 datasets. Interestingly, it also outperforms more powerful VQA models such as Multimodal Compact Bilinear Pooling (MCB) in several cases. GVQA offers strengths complementary to SAN when trained and evaluated on the original VQA v1 and VQA v2 datasets. Finally, GVQA is more transparent and interpretable than existing VQA models.

Most industrial recommender systems rely on the popular collaborative filtering (CF) technique for providing personalized recommendations to its users. However, the very nature of CF is adversarial to the idea of user privacy, because users need to share their preferences with others in order to be grouped with like-minded people and receive accurate recommendations. Prior related work have proposed to preserve user privacy in a CF framework through different means like (i) random data obfuscation using differential privacy techniques, (ii) relying on decentralized trusted peer networks, or (iii) by adopting secured cryptographic strategies. While these approaches have been successful inasmuch as they concealed user preference information to some extent from a centralized recommender system, they have also, nevertheless, incurred significant tradeoffs in terms of privacy, scalability, and accuracy. They are also vulnerable to privacy breaches by malicious actors. In light of these observations, we propose a novel selective privacy preserving (SP2) paradigm that allows users to custom define the scope and extent of their individual privacies, by marking their personal ratings as either public (which can be shared) or private (which are never shared and stored only on the user device). Our SP2 framework works in two steps: (i) First, it builds an initial recommendation model based on the sum of all public ratings that have been shared by users and (ii) then, this public model is fine-tuned on each user's device based on the user private ratings, thus eventually learning a more accurate model. Furthermore, in this work, we introduce three different algorithms for implementing an end-to-end SP2 framework that can scale effectively from thousands to hundreds of millions of items.

Current deep learning based text classification methods are limited by their ability to achieve fast learning and generalization when the data is scarce. We address this problem by integrating a meta-learning procedure that uses the knowledge learned across many tasks as an inductive bias towards better natural language understanding. Based on the Model-Agnostic Meta-Learning framework (MAML), we introduce the Attentive Task-Agnostic Meta-Learning (ATAML) algorithm for text classification. The essential difference between MAML and ATAML is in the separation of task-agnostic representation learning and task-specific attentive adaptation. The proposed ATAML is designed to encourage task-agnostic representation learning by way of task-agnostic parameterization and facilitate task-specific adaptation via attention mechanisms. We provide evidence to show that the attention mechanism in ATAML has a synergistic effect on learning performance. In comparisons with models trained from random initialization, pretrained models and meta trained MAML, our proposed ATAML method generalizes better on single-label and multi-label classification tasks in miniRCV1 and miniReuters-21578 datasets.

This paper studies a Nystr\"om type subsampling approach to large kernel learning methods in the misspecified case, where the target function is not assumed to belong to the reproducing kernel Hilbert space generated by the underlying kernel. This case is less understood, in spite of its practical importance. To model such a case, the smoothness of target functions is described in terms of general source conditions. It is surprising that almost for the whole range of the source conditions, describing the misspecified case, the corresponding learning rate bounds can be achieved with just one value of the regularization parameter. This observation allows a formulation of mild conditions under which the plain Nystr\"om subsampling can be realized with subquadratic cost maintaining the guaranteed learning rates.

In recent years, there has been a surge of interest in developing deep learning methods for non-Euclidean structured data such as graphs. In this paper, we propose Dual-Primal Graph CNN, a graph convolutional architecture that alternates convolution-like operations on the graph and its dual. Our approach allows to learn both vertex- and edge features and generalizes the previous graph attention (GAT) model. We provide extensive experimental validation showing state-of-the-art results on a variety of tasks tested on established graph benchmarks, including CORA and Citeseer citation networks as well as MovieLens, Flixter, Douban and Yahoo Music graph-guided recommender systems.

Schizophrenia, a mental disorder that is characterized by abnormal social behavior and failure to distinguish one's own thoughts and ideas from reality, has been associated with structural abnormalities in the architecture of functional brain networks. Using various methods from network analysis, we examine the effect of two classical therapeutic antipsychotics --- Aripiprazole and Sulpiride --- on the structure of functional brain networks of healthy controls and patients who have been diagnosed with schizophrenia. We compare the community structures of functional brain networks of different individuals using mesoscopic response functions, which measure how community structure changes across different scales of a network. We are able to do a reasonably good job of distinguishing patients from controls, and we are most successful at this task on people who have been treated with Aripiprazole. We demonstrate that this increased separation between patients and controls is related only to a change in the control group, as the functional brain networks of the patient group appear to be predominantly unaffected by this drug. This suggests that Aripiprazole has a significant and measurable effect on community structure in healthy individuals but not in individuals who are diagnosed with schizophrenia. In contrast, we find for individuals are given the drug Sulpiride that it is more difficult to separate the networks of patients from those of controls. Overall, we observe differences in the effects of the drugs (and a placebo) on community structure in patients and controls and also that this effect differs across groups. We thereby demonstrate that different types of antipsychotic drugs selectively affect mesoscale structures of brain networks, providing support that mesoscale structures such as communities are meaningful functional units in the brain.

We provide a new understanding of the fundamental nature of adversarially robust classifiers and how they differ from standard models. In particular, we show that there provably exists a trade-off between the standard accuracy of a model and its robustness to adversarial perturbations. We demonstrate an intriguing phenomenon at the root of this tension: a certain dichotomy between "robust" and "non-robust" features. We show that while robustness comes at a price, it also has some surprising benefits. Robust models turn out to have interpretable gradients and feature representations that align unusually well with salient data characteristics. In fact, they yield striking feature interpolations that have thus far been possible to obtain only using generative models such as GANs.

Neural attention (NA) is an effective mechanism for inferring complex structural data dependencies that span long temporal horizons. As a consequence, it has become a key component of sequence-to-sequence models that yield state-of-the-art performance in as hard tasks as abstractive document summarization (ADS), machine translation (MT), and video captioning (VC). NA mechanisms perform inference of context vectors; these constitute weighted sums of deterministic input sequence encodings, adaptively sourced over long temporal horizons. However, recent work in the field of amortized variational inference (AVI) has shown that it is often useful to treat the representations generated by deep networks as latent random variables. This allows for the models to better explore the space of possible representations. Based on this motivation, in this work we introduce a novel regard towards a popular NA mechanism, namely soft-attention (SA). Our approach treats the context vectors generated by SA models as latent variables, the posteriors of which are inferred by employing AVI. Both the means and the covariance matrices of the inferred posteriors are parameterized via deep network mechanisms similar to those employed in the context of standard SA. To illustrate our method, we implement it in the context of popular sequence-to-sequence model variants with SA. We conduct an extensive experimental evaluation using challenging ADS, VC, and MT benchmarks, and show how our approach compares to the baselines.

Digital texts have become an increasingly important source of data for social studies. However, textual data from open platforms are vulnerable to manipulation (e.g., censorship and information inflation), often leading to bias in subsequent empirical analysis. This paper investigates the problem of data distortion in text classification when controlling type I error (a relevant textual message is classified as irrelevant) is the priority. The default classical classification paradigm that minimizes the overall classification error can yield an undesirably large type I error, and data distortion exacerbates this situation. As a solution, we propose the Neyman-Pearson (NP) classification paradigm which minimizes type II error under a user-specified type I error constraint. Theoretically, we show that while the classical oracle (i.e., optimal classifier) cannot be recovered under unknown data distortion even if one has the entire post-distortion population, the NP oracle is unaffected by data distortion and can be recovered under the same condition. Empirically, we illustrate the advantage of NP classification methods in a case study that classifies posts about strikes and corruption published on a leading Chinese blogging platform.

In classification applications such as severe disease diagnosis and fraud detection, people have clear priorities over the two types of classification errors. For instance, diagnosing a patient with cancer to be healthy may lead to loss of life, which incurs a much higher cost than the other way around. The classical binary classification paradigm does not take into account such priorities, as it aims to minimize the overall classification error. In contrast, the Neyman-Pearson (NP) paradigm seeks classifiers with a minimal type II error while having the prioritized type I error constrained under a user-specified level, addressing asymmetric type I/II error priorities in the previously mentioned scenarios. Despite recent advances in the NP classification literature, two essential issues pose challenges: i) current theoretical framework assumes bounded feature support, which does not admit parametric settings; ii) in practice, existing NP classifiers involve splitting class 0 samples into two parts using a pre-fixed split proportion. To address the first challenge, we present NP-sLDA that adapts the popular sparse linear discriminant analysis (sLDA, Mai et al. (2012)) to the NP paradigm. On the theoretical front, this is the first theoretically justified NP classifier that takes parametric assumptions and unbounded feature support. We formulate a new conditional margin assumption and a new conditional detection condition to accommodate unbounded feature support and show that NP-sLDA satisfies the NP oracle inequalities. Numerical results show that NP-sLDA is a valuable addition to the existing NP classifiers. To address the second challenge, we construct a general data-adaptive sample splitting scheme that improves the classification performance upon the default half-half class 0 split used in Tong et al. (2018).