When watching videos, the occurrence of a visual event is often accompanied by an audio event, e.g., the voice of lip motion, the music of playing instruments. There is an underlying correlation between audio and visual events, which can be utilized as free supervised information to train a neural network by solving the pretext task of audio-visual synchronization. In this paper, we propose a novel self-supervised framework with co-attention mechanism to learn generic cross-modal representations from unlabelled videos in the wild, and further benefit downstream tasks. Specifically, we explore three different co-attention modules to focus on discriminative visual regions correlated to the sounds and introduce the interactions between them. Experiments show that our model achieves state-of-the-art performance on the pretext task while having fewer parameters compared with existing methods. To further evaluate the generalizability and transferability of our approach, we apply the pre-trained model on two downstream tasks, i.e., sound source localization and action recognition. Extensive experiments demonstrate that our model provides competitive results with other self-supervised methods, and also indicate that our approach can tackle the challenging scenes which contain multiple sound sources.

We present a data-driven approach using word embeddings to discover and categorise language biases on the discussion platform Reddit. As spaces for isolated user communities, platforms such as Reddit are increasingly connected to issues of racism, sexism and other forms of discrimination. Hence, there is a need to monitor the language of these groups. One of the most promising AI approaches to trace linguistic biases in large textual datasets involves word embeddings, which transform text into high-dimensional dense vectors and capture semantic relations between words. Yet, previous studies require predefined sets of potential biases to study, e.g., whether gender is more or less associated with particular types of jobs. This makes these approaches unfit to deal with smaller and community-centric datasets such as those on Reddit, which contain smaller vocabularies and slang, as well as biases that may be particular to that community. This paper proposes a data-driven approach to automatically discover language biases encoded in the vocabulary of online discourse communities on Reddit. In our approach, protected attributes are connected to evaluative words found in the data, which are then categorised through a semantic analysis system. We verify the effectiveness of our method by comparing the biases we discover in the Google News dataset with those found in previous literature. We then successfully discover gender bias, religion bias, and ethnic bias in different Reddit communities. We conclude by discussing potential application scenarios and limitations of this data-driven bias discovery method.

A solid methodology to understand human perception and preferences in human-robot interaction (HRI) is crucial in designing real-world HRI. Social cognition posits that the dimensions Warmth and Competence are central and universal dimensions characterizing other humans. The Robotic Social Attribute Scale (RoSAS) proposes items for those dimensions suitable for HRI and validated them in a visual observation study. In this paper we complement the validation by showing the usability of these dimensions in a behavior based, physical HRI study with a fully autonomous robot. We compare the findings with the popular Godspeed dimensions Animacy, Anthropomorphism, Likeability, Perceived Intelligence and Perceived Safety. We found that Warmth and Competence, among all RoSAS and Godspeed dimensions, are the most important predictors for human preferences between different robot behaviors. This predictive power holds even when there is no clear consensus preference or significant factor difference between conditions.

Clustering is essential for exploratory data mining, data structure analysis and a common technique for statistical data analysis. It is widely used in many fields, including machine learning, pattern recognition, image analysis, information retrieval, bioinformatics, data compression, and computer graphics. Many clustering approaches exist with different intrinsic notions of what a cluster is. In the standard setup, the goal is to group objects into subsets, known as clusters, such that objects within a given cluster are more related to one another than the ones from a different cluster. Clustering is already quite well-known (see [4, 27] and references therein) conversely to Extreme Value Theory (EVT) which is a newer field in the machine learning community that has been used in anomaly detection [14, 28, 45, 51], classification [31, 32, 54] or clustering [10, 12, 13, 33] when dedicated to the most extreme regions of the sample space.

We investigate optimal posteriors for recently introduced \cite{begin2016pac} chi-squared divergence based PAC-Bayesian bounds in terms of nature of their distribution, scalability of computations, and test set performance. For a finite classifier set, we deduce bounds for three distance functions: KL-divergence, linear and squared distances. Optimal posterior weights are proportional to deviations of empirical risks, usually with subset support. For uniform prior, it is sufficient to search among posteriors on classifier subsets ordered by these risks. We show the bound minimization for linear distance as a convex program and obtain a closed-form expression for its optimal posterior. Whereas that for squared distance is a quasi-convex program under a specific condition, and the one for KL-divergence is non-convex optimization (a difference of convex functions). To compute such optimal posteriors, we derive fast converging fixed point (FP) equations. We apply these approaches to a finite set of SVM regularization parameter values to yield stochastic SVMs with tight bounds. We perform a comprehensive performance comparison between our optimal posteriors and known KL-divergence based posteriors on a variety of UCI datasets with varying ranges and variances in risk values, etc. Chi-squared divergence based posteriors have weaker bounds and worse test errors, hinting at an underlying regularization by KL-divergence based posteriors. Our study highlights the impact of divergence function on the performance of PAC-Bayesian classifiers. We compare our stochastic classifiers with cross-validation based deterministic classifier. The latter has better test errors, but ours is more sample robust, has quantifiable generalization guarantees, and is computationally much faster.

Networks can describe the structure of a wide variety of complex systems by specifying how pairs of nodes interact. This choice of representation is flexible, but not necessarily appropriate when joint interactions between groups of nodes are needed to explain empirical phenomena. Networks remain the de facto standard, however, as relational datasets often fail to include higher-order interactions. Here, we introduce a Bayesian approach to reconstruct these missing higher-order interactions, from pairwise network data. Our method is based on the principle of parsimony and only includes higher-order structures when there is sufficient statistical evidence for them.

The reliability of machine learning systems critically assumes that the associations between features and labels remain similar between training and test distributions. However, unmeasured variables, such as confounders, break this assumption---useful correlations between features and labels at training time can become useless or even harmful at test time. For example, high obesity is generally predictive for heart disease, but this relation may not hold for smokers who generally have lower rates of obesity and higher rates of heart disease. We present a framework for making models robust to spurious correlations by leveraging humans' common sense knowledge of causality. Specifically, we use human annotation to augment each training example with a potential unmeasured variable (i.e. an underweight patient with heart disease may be a smoker), reducing the problem to a covariate shift problem. We then introduce a new distributionally robust optimization objective over unmeasured variables (UV-DRO) to control the worst-case loss over possible test-time shifts. Empirically, we show improvements of 5-10% on a digit recognition task confounded by rotation, and 1.5-5% on the task of analyzing NYPD Police Stops confounded by location.

While nowadays most gradient-based optimization methods focus on exploring the high-dimensional geometric features, the random error accumulated in a stochastic version of any algorithm implementation has not been stressed yet. In this work, we propose a universal principle which reduces the random error accumulation by exploiting statistic information hidden in mini-batch gradients. This is achieved by regularizing the learning-rate according to mini-batch variances. Due to the complementarity of our perspective, this regularization could provide a further improvement for stochastic implementation of generic 1st order approaches. With empirical results, we demonstrated the variance regularization could speed up the convergence as well as stabilize the stochastic optimization.

This article is motivated by the problem of selecting the functional form of trees and ensemble size in gradient tree boosting (Friedman, 2001; Mason et al., 2000). Gradient tree boosting (GTB) has become extremely popular in recent years, both in academia and industry: At present, an increase in the size of datasets, both in the number of observations and the richness of the data, or number of features, is seen. This, coupled with an exponential increase in computational power and a growing revelation and acceptance for data-driven decisions in the industry makes for an increasing interest in statistical learning (Hastie et al., 2001). For these new datasets, standard statistical methods such as generalized linear models (McCullagh and Nelder, 1989) that have a fixed learning rate due to their constrained functional form with bounded complexity, struggle in terms of predictive power, as they stop learning at certain information thresholds. The interest is therefore geared towards more flexible approaches such as ensembles of learners.

In domains where data are sensitive or private, there is great value in methods that can learn in a distributed manner without the data ever leaving the local devices. In light of this need, federated learning has emerged as a popular training paradigm. However, many federated learning approaches trade transmitting data for communicating updated weight parameters for each local device. Therefore, a successful breach that would have otherwise directly compromised the data instead grants whitebox access to the local model, which opens the door to a number of attacks, including exposing the very data federated learning seeks to protect. Additionally, in distributed scenarios, individual client devices commonly exhibit high statistical heterogeneity. Many common federated approaches learn a single global model; while this may do well on average, performance degrades when the i.i.d. assumption is violated, underfitting individuals further from the mean, and raising questions of fairness. To address these issues, we propose Weight Anonymized Factorization for Federated Learning (WAFFLe), an approach that combines the Indian Buffet Process with a shared dictionary of weight factors for neural networks. Experiments on MNIST, FashionMNIST, and CIFAR-10 demonstrate WAFFLe's significant improvement to local test performance and fairness while simultaneously providing an extra layer of security.