Typically, machine learning models are trained and evaluated without making any distinction between users (e.g, using traditional hold-out and cross-validation). However, this produces inaccurate performance metrics estimates in multi-user settings. That is, situations where the data were collected by multiple users with different characteristics (e.g., age, gender, height, etc.) which is very common in user computer interaction and medical applications. For these types of scenarios model evaluation strategies that provide better performance estimates have been proposed such as mixed, user-independent, user-dependent, and user-adaptive models. Although those strategies are better suited for multi-user systems, they are typically assessed with respect to performance metrics that capture the overall behavior of the models and do not provide any performance guarantees for individual predictions nor they provide any feedback about the predictions' uncertainty. In order to overcome those limitations, in this work we evaluated the conformal prediction framework in several multi-user settings. Conformal prediction is a model agnostic method that provides confidence guarantees on the predictions, thus, increasing the trustworthiness and robustness of the models. We conducted extensive experiments using different evaluation strategies and found significant differences in terms of conformal performance measures. We also proposed several visualizations based on matrices, graphs, and charts that capture different aspects of the resulting prediction sets.

Large-scale automatic speech translation systems today lack key features that help machine-mediated communication feel seamless when compared to human-to-human dialogue. In this work, we introduce a family of models that enable end-to-end expressive and multilingual translations in a streaming fashion. First, we contribute an improved version of the massively multilingual and multimodal SeamlessM4T model-SeamlessM4T v2. This newer model, incorporating an updated UnitY2 framework, was trained on more low-resource language data. SeamlessM4T v2 provides the foundation on which our next two models are initiated. SeamlessExpressive enables translation that preserves vocal styles and prosody. Compared to previous efforts in expressive speech research, our work addresses certain underexplored aspects of prosody, such as speech rate and pauses, while also preserving the style of one's voice. As for SeamlessStreaming, our model leverages the Efficient Monotonic Multihead Attention mechanism to generate low-latency target translations without waiting for complete source utterances. As the first of its kind, SeamlessStreaming enables simultaneous speech-to-speech/text translation for multiple source and target languages. To ensure that our models can be used safely and responsibly, we implemented the first known red-teaming effort for multimodal machine translation, a system for the detection and mitigation of added toxicity, a systematic evaluation of gender bias, and an inaudible localized watermarking mechanism designed to dampen the impact of deepfakes. Consequently, we bring major components from SeamlessExpressive and SeamlessStreaming together to form Seamless, the first publicly available system that unlocks expressive cross-lingual communication in real-time. The contributions to this work are publicly released and accessible at https://github.com/facebookresearch/seamless_communication

Recently, diffusion models have become popular tools for image synthesis because of their high-quality outputs. However, like other large-scale models, they may leak private information about their training data. Here, we demonstrate a privacy vulnerability of diffusion models through a \emph{membership inference (MI) attack}, which aims to identify whether a target example belongs to the training set when given the trained diffusion model. Our proposed MI attack learns quantile regression models that predict (a quantile of) the distribution of reconstruction loss on examples not used in training. This allows us to define a granular hypothesis test for determining the membership of a point in the training set, based on thresholding the reconstruction loss of that point using a custom threshold tailored to the example. We also provide a simple bootstrap technique that takes a majority membership prediction over ``a bag of weak attackers'' which improves the accuracy over individual quantile regression models. We show that our attack outperforms the prior state-of-the-art attack while being substantially less computationally expensive -- prior attacks required training multiple ``shadow models'' with the same architecture as the model under attack, whereas our attack requires training only much smaller models.

Data collected about individuals is regularly used to make decisions that impact those same individuals. For example, census statistics have important implications for all aspects of daily life, including the allocation of political power, the distribution of federal funds, and research in economics and social sciences. In banking industries, machine learning (ML) models leverage data to proactively monitor customer behavior, reduce the likelihood of false positives, and prevent fraud. In these settings, there is a tension between the need for accurate systems, in which individuals receive what they deserve, and the need to protect individuals from improper disclosure of their sensitive information. Differential privacy (DP) [23] is now widely recognized as the gold standard for providing formal guarantees on the privacy level achieved by an algorithm. However, central DP can only be used on the assumption of a trustworthy server. Local DP (LDP) [32] is a variant that achieves privacy guarantees for each user locally with no assumptions on third-party servers. In other words, LDP ensures that each user's data is locally obfuscated first on the client-side and then sent to the server-side, thus protecting data from privacy leaks on both the client-side and the server-side. Many Big tech companies have deployed LDP-based algorithms to use in their industrial products (e.g., Google Chrome [24] and Apple iOS [4]).

We reveal a one-class homophily phenomenon, which is one prevalent property we find empirically in real-world graph anomaly detection (GAD) datasets, i.e., normal nodes tend to have strong connection/affinity with each other, while the homophily in abnormal nodes is significantly weaker than normal nodes. However, this anomaly-discriminative property is ignored by existing GAD methods that are typically built using a conventional anomaly detection objective, such as data reconstruction. In this work, we explore this property to introduce a novel unsupervised anomaly scoring measure for GAD - local node affinity - that assigns a larger anomaly score to nodes that are less affiliated with their neighbors, with the affinity defined as similarity on node attributes/representations. We further propose Truncated Affinity Maximization (TAM) that learns tailored node representations for our anomaly measure by maximizing the local affinity of nodes to their neighbors. Optimizing on the original graph structure can be biased by nonhomophily edges (i.e., edges connecting normal and abnormal nodes). Thus, TAM is instead optimized on truncated graphs where non-homophily edges are removed iteratively to mitigate this bias. The learned representations result in significantly stronger local affinity for normal nodes than abnormal nodes. Extensive empirical results on 10 real-world GAD datasets show that TAM substantially outperforms seven competing models, achieving over 10% increase in AUROC/AUPRC compared to the best contenders on challenging datasets.

Detecting Parkinson's Disease in its early stages using EEG data presents a significant challenge. This paper introduces a novel approach, representing EEG data as a 15-variate series of bandpower and peak frequency values/coefficients. The hypothesis is that this representation captures essential information from the noisy EEG signal, improving disease detection. Statistical features extracted from this representation are utilised as input for interpretable machine learning models, specifically Decision Tree and AdaBoost classifiers. Our classification pipeline is deployed within our proposed framework which enables high-importance data types and brain regions for classification to be identified. Interestingly, our analysis reveals that while there is no significant regional importance, the N1 sleep data type exhibits statistically significant predictive power (p < 0.01) for early-stage Parkinson's Disease classification. AdaBoost classifiers trained on the N1 data type consistently outperform baseline models, achieving over 80% accuracy and recall. Our classification pipeline statistically significantly outperforms baseline models indicating that the model has acquired useful information. Paired with the interpretability (ability to view feature importance's) of our pipeline this enables us to generate meaningful insights into the classification of early stage Parkinson's with our N1 models. In Future, these models could be deployed in the real world - the results presented in this paper indicate that more than 3 in 4 early-stage Parkinson's cases would be captured with our pipeline.

Estimating the prevalence of a medical condition, or the proportion of the population in which it occurs, is a fundamental problem in healthcare and public health. Accurate estimates of the relative prevalence across groups -- capturing, for example, that a condition affects women more frequently than men -- facilitate effective and equitable health policy which prioritizes groups who are disproportionately affected by a condition. However, it is difficult to estimate relative prevalence when a medical condition is underreported. In this work, we provide a method for accurately estimating the relative prevalence of underreported medical conditions, building upon the positive unlabeled learning framework. We show that under the commonly made covariate shift assumption -- i.e., that the probability of having a disease conditional on symptoms remains constant across groups -- we can recover the relative prevalence, even without restrictive assumptions commonly made in positive unlabeled learning and even if it is impossible to recover the absolute prevalence. We conduct experiments on synthetic and real health data which demonstrate our method's ability to recover the relative prevalence more accurately than do baselines, and demonstrate the method's robustness to plausible violations of the covariate shift assumption. We conclude by illustrating the applicability of our method to case studies of intimate partner violence and hate speech.

Supervised imitation learning, also known as behavioral cloning, suffers from distribution drift leading to failures during policy execution. One approach to mitigate this issue is to allow an expert to correct the agent's actions during task execution, based on the expert's determination that the agent has reached a `point of no return.' The agent's policy is then retrained using this new corrective data. This approach alone can enable high-performance agents to be learned, but at a substantial cost: the expert must vigilantly observe execution until the policy reaches a specified level of success, and even at that point, there is no guarantee that the policy will always succeed. To address these limitations, we present FIRE (Failure Identification to Reduce Expert Burden in intervention-based learning), a system that can predict when a running policy will fail, halt its execution, and request a correction from the expert. Unlike existing approaches that learn only from expert data, our approach learns from both expert and non-expert data, akin to adversarial learning. We demonstrate experimentally for a series of challenging manipulation tasks that our method is able to recognize state-action pairs that lead to failures. This permits seamless integration into an intervention-based learning system, where we show an order-of-magnitude gain in sample efficiency compared with a state-of-the-art inverse reinforcement learning method and dramatically improved performance over an equivalent amount of data learned with behavioral cloning.

Multi-source domain adaptation (DA) aims at leveraging information from more than one source domain to make predictions in a target domain, where different domains may have different data distributions. Most existing methods for multi-source DA focus on classification problems while there is only limited investigation in the regression settings. In this paper, we fill in this gap through a two-step procedure. First, we extend a flexible single-source DA algorithm for classification through outcome-coarsening to enable its application to regression problems. We then augment our single-source DA algorithm for regression with ensemble learning to achieve multi-source DA. We consider three learning paradigms in the ensemble algorithm, which combines linearly the target-adapted learners trained with each source domain: (i) a multi-source stacking algorithm to obtain the ensemble weights; (ii) a similarity-based weighting where the weights reflect the quality of DA of each target-adapted learner; and (iii) a combination of the stacking and similarity weights. We illustrate the performance of our algorithms with simulations and a data application where the goal is to predict High-density lipoprotein (HDL) cholesterol levels using gut microbiome. We observe a consistent improvement in prediction performance of our multi-source DA algorithm over the routinely used methods in all these scenarios.

The is no other model or hypothesis verification tool in Bayesian statistics that is as widely used as the Bayes factor. We focus on generative models that are likelihood-free and, therefore, render the computation of Bayes factors (marginal likelihood ratios) far from obvious. We propose a deep learning estimator of the Bayes factor based on simulated data from two competing models using the likelihood ratio trick. This estimator is devoid of summary statistics and obviates some of the difficulties with ABC model choice. We establish sufficient conditions for consistency of our Deep Bayes Factor estimator as well as its consistency as a model selection tool. We investigate the performance of our estimator on various examples using a wide range of quality metrics related to estimation and model decision accuracy. After training, our deep learning approach enables rapid evaluations of the Bayes factor estimator at any fictional data arriving from either hypothesized model, not just the observed data $Y_0$. This allows us to inspect entire Bayes factor distributions under the two models and to quantify the relative location of the Bayes factor evaluated at $Y_0$ in light of these distributions. Such tail area evaluations are not possible for Bayes factor estimators tailored to $Y_0$. We find the performance of our Deep Bayes Factors competitive with existing MCMC techniques that require the knowledge of the likelihood function. We also consider variants for posterior or intrinsic Bayes factors estimation. We demonstrate the usefulness of our approach on a relatively high-dimensional real data example about determining cognitive biases.