Large Language Models (LLMs) are prone to generating content that exhibits gender biases, raising significant ethical concerns. Alignment, the process of fine-tuning LLMs to better align with desired behaviors, is recognized as an effective approach to mitigate gender biases. Although proprietary LLMs have made significant strides in mitigating gender bias, their alignment datasets are not publicly available. The commonly used and publicly available alignment dataset, HH-RLHF, still exhibits gender bias to some extent. There is a lack of publicly available alignment datasets specifically designed to address gender bias. Hence, we developed a new dataset named GenderAlign, aiming at mitigating a comprehensive set of gender biases in LLMs. This dataset comprises 8k single-turn dialogues, each paired with a “chosen” and a “rejected” response. Compared to the “rejected” responses, the “chosen” responses demonstrate lower levels of gender bias and higher quality. Furthermore, we categorized the gender biases in the “rejected” responses of GenderAlign into 4 principal categories. The experimental results show the effectiveness of GenderAlign in reducing gender bias in LLMs.

Traditional large language models (LLMs) leverage extensive text corpora but lack access to acoustic and para-linguistic cues present in speech. There is a growing interest in enhancing text-based models with audio information. However, current models often require an aligned audio-text dataset which is frequently much smaller than typical language model training corpora. Moreover, these models often require both text and audio streams during inference/testing. In this study, we introduce a novel two-stage knowledge distillation (KD) approach that enables language models to (a) incorporate rich acoustic and paralinguistic information from speech, (b) utilize text corpora comparable in size to typical language model training data, and (c) support text-only analysis without requiring an audio stream during inference/testing. Specifically, we employ a pre-trained speech embedding teacher model (OpenAI Whisper) to train a Teacher Assistant (TA) model on an aligned audio-text dataset in the first stage. In the second stage, the TA’s knowledge is transferred to a student language model trained on a conventional text dataset. Thus, our two-stage KD method leverages both the acoustic and paralinguistic cues in the aligned audio-text data and the nuanced linguistic knowledge in a large text-only dataset. Based on our evaluation, this DoubleDistillation system consistently outperforms traditional LLMs in 15 informal text understanding tasks.

Floods are highly destructive natural disasters that result in significant economic losses and endanger human and wildlife lives. Efficiently monitoring Flooded areas through the utilization of deep learning models can contribute to mitigating these risks. This study focuses on the deployment of deep learning models specifically designed for classifying flooded and non-flooded in UAV images. In consideration of computational costs, we propose modified version of ResNet50 called Flood-ResNet50. By incorporating additional layers and leveraging transfer learning techniques, Flood-ResNet50 achieves comparable performance to larger models like VGG16/19, AlexNet, DenseNet161, EfficientNetB7, Swin(small), and vision transformer. Experimental results demonstrate that the proposed modification of ResNet50, incorporating additional layers, achieves a classification accuracy of 96.43%, F1 score of 86.36%, Recall of 81.11%, Precision of 92.41 %, model size 98MB and FLOPs 4.3 billions for the FloodNet dataset. When deployed on edge devices such as the Jetson Nano, our model demonstrates faster inference speed (820 ms), higher throughput (39.02 fps), and lower average power consumption (6.9 W) compared to larger ResNet101 and ResNet152 models.

Currently, there is a surge of interest in fair Artificial Intelligence (AI) and Machine Learning (ML) research which aims to mitigate discriminatory bias in AI algorithms, e.g., along lines of gender, age, and race. While most research in this domain focuses on developing fair AI algorithms, in this work, we examine the challenges which arise when humans and fair AI interact. Our results show that due to an apparent conflict between human preferences and fairness, a fair AI algorithm on its own may be insufficient to achieve its intended results in the real world. Using college major recommendation as a case study, we build a fair AI recommender by employing gender debiasing machine learning techniques. Our offline evaluation showed that the debiased recommender makes fairer career recommendations without sacrificing its accuracy in prediction. Nevertheless, an online user study of more than 200 college students revealed that participants on average prefer the original biased system over the debiased system. Specifically, we found that perceived gender disparity is a determining factor for the acceptance of a recommendation. In other words, we cannot fully address the gender bias issue in AI recommendations without addressing the gender bias in humans. We conducted a follow-up survey to gain additional insights into the effectiveness of various design options that can help participants to overcome their own biases. Our results suggest that making fair AI explainable is crucial for increasing its adoption in the real world.

The Arctic is a region with unique climate features, motivating new AI methodologies to study it. Unfortunately, Arctic sea ice has seen a continuous decline since 1979. This not only poses a significant threat to Arctic wildlife and surrounding coastal communities but is also adversely affecting the global climate patterns. To study the potential of AI in tackling climate change, we analyze the performance of four probabilistic machine learning methods in forecasting sea-ice extent for lead times of up to 6 months, further comparing them with traditional machine learning methods. Our comparative analysis shows that Gaussian Process Regression is a good fit to predict sea-ice extent for longer lead times with lowest RMSE score.

Currently, there is a surge of interest in fair Artificial Intelligence (AI) and Machine Learning (ML) research which aims to mitigate discriminatory bias in AI algorithms, e.g. along lines of gender, age, and race. While most research in this domain focuses on developing fair AI algorithms, in this work, we examine the challenges which arise when human- fair-AI interact. Our results show that due to an apparent conflict between human preferences and fairness, a fair AI algorithm on its own may be insufficient to achieve its intended results in the real world. Using college major recommendation as a case study, we build a fair AI recommender by employing gender debiasing machine learning techniques. Our offline evaluation showed that the debiased recommender makes fairer and more accurate college major recommendations. Nevertheless, an online user study of more than 200 college students revealed that participants on average prefer the original biased system over the debiased system. Specifically, we found that the perceived gender disparity associated with a college major is a determining factor for the acceptance of a recommendation. In other words, our results demonstrate we cannot fully address the gender bias issue in AI recommendations without addressing the gender bias in humans. They also highlight the urgent need to extend the current scope of fair AI research from narrowly focusing on debiasing AI algorithms to including new persuasion and bias explanation technologies in order to achieve intended societal impacts.

Abstract We propose a method that uses neural embeddings to improve the performance of any given LDA-style topic model. Our method, called neural embedding allocation (NEA), deconstructs topic models (LDA or otherwise) into interpretable vector-space embeddings of words, topics, documents, authors, and so on, by learning neural embeddings to mimic the topic model. We demonstrate that NEA improves coherence scores of the original topic model by smoothing out the noisy topics when the number of topics is large. Furthermore, we show NEA’s effectiveness and generality in deconstructing and smoothing LDA, author-topic models, and the recent mixed membership skip-gram topic model and achieve better performance with the embeddings compared to several state-of-the-art models.

There is growing awareness that AI and machine learning systems can in some cases learn to behave in unfair and discriminatory ways with harmful consequences. However, despite an enormous amount of research, techniques for ensuring AI fairness have yet to see widespread deployment in real systems. One of the main barriers is the conventional wisdom that fairness brings a cost in predictive performance metrics such as accuracy which could affect an organization's bottom-line. In this paper we take a closer look at this concern. Clearly fairness/performance trade-offs exist, but are they inevitable? In contrast to the conventional wisdom, we find that it is frequently possible, indeed straightforward, to improve on a trained model's fairness without sacrificing predictive performance. We systematically study the behavior of fair learning algorithms on a range of benchmark datasets, showing that it is possible to improve fairness to some degree with no loss (or even an improvement) in predictive performance via a sensible hyper-parameter selection strategy. Our results reveal a pathway toward increasing the deployment of fair AI methods, with potentially substantial positive real-world impacts.

Healthcare programs such as Medicaid provide crucial services to vulnerable populations, but due to limited resources, many of the individuals who need these services the most languish on waiting lists.Survival models, e.g. the Cox proportional hazards model, can potentially improve this situation by predicting individuals' levels of need, which can then be used to prioritize the waiting lists.Providing care to those in need can prevent institutionalization for those individuals, which both improves quality of life and reduces overall costs.While the benefits of such an approach are clear, care must be taken to ensure that the prioritization process is fair, and does not reinforce harmful systemic bias.We develop multiple fairness definitions and corresponding fair learning algorithms for survival models to ensure equitable allocation of healthcare resources.We demonstrate the utility of our methods in terms of fairness and predictive accuracy on three publicly available survival datasets.

A growing proportion of human interactions are digitized on social media platforms and subjected to algorithmic decision-making, and it has become increasingly important to ensure fair treatment from these algorithms. In this work, we investigate gender bias in collaborative-filtering recommender systems trained on social media data. We develop neural fair collaborative filtering (NFCF), a practical framework for mitigating gender bias in recommending career-related sensitive items (e.g. jobs, academic concentrations, or courses of study) using a pre-training and fine-tuning approach to neural collaborative filtering, augmented with bias correction techniques. We show the utility of our methods for gender de-biased career and college major recommendations on the MovieLens dataset and a Facebook dataset, respectively, and achieve better performance and fairer behavior than several state-of-the-art models.

We propose differential fairness, a multi-attribute definition of fairness in machine learning which is informed by intersectionality, a critical lens arising from the humanities literature, leveraging connections between differential privacy and legal notions of fairness. We show that our criterion behaves sensibly for any subset of the set of protected attributes, and we prove economic, privacy, and generalization guarantees. We provide a learning algorithm which respects our differential fairness criterion. Experiments on the COMPAS criminal recidivism dataset and census data demonstrate the utility of our methods.

Many applications of Bayesian data analysis involve sensitive information such as personal documents or medical records, motivating methods which ensure that privacy is protected. We introduce a general privacy-preserving framework for Variational Bayes (VB), a widely used optimization-based Bayesian inference method. Our framework respects differential privacy, the gold-standard privacy criterion. The iterative nature of variational Bayes presents a challenge since iterations increase the amount of noise needed to ensure privacy. We overcome this by combining: (1) an improved composition method, called the moments accountant, and (2) the privacy amplification effect of subsampling mini-batches from large-scale data in stochastic learning. We empirically demonstrate the effectiveness of our method on LDA topic models, evaluated on Wikipedia. In the full paper we extend our method to a broad class of models, including Bayesian logistic regression and sigmoid belief networks.

Intersectionality is a framework that analyzes how interlocking systems of power and oppression affect individuals along overlapping dimensions including race, gender, sexual orientation, class, and disability. Intersectionality theory therefore implies it is important that fairness in artificial intelligence systems be protected with regard to multi-dimensional protected attributes. However, the measurement of fairness becomes statistically challenging in the multi-dimensional setting due to data sparsity, which increases rapidly in the number of dimensions, and in the values per dimension. We present a Bayesian probabilistic modeling approach for the reliable, data-efficient estimation of fairness with multidimensional protected attributes, which we apply to two existing intersectional fairness metrics. Experimental results on census data and the COMPAS criminal justice recidivism dataset demonstrate the utility of our methodology, and show that Bayesian methods are valuable for the modeling and measurement of fairness in intersectional contexts.

Many applications of Bayesian data analysis involve sensitive information such as personal documents or medical records, motivating methods which ensure that privacy is protected. We introduce a general privacy-preserving framework for Variational Bayes (VB), a widely used optimization-based Bayesian inference method. Our framework respects differential privacy, the gold-standard privacy criterion, and encompasses a large class of probabilistic models, called the Conjugate Exponential (CE) family. We observe that we can straightforwardly privatise VB’s approximate posterior distributions for models in the CE family, by perturbing the expected sufficient statistics of the complete-data likelihood. For a broadly-used class of non-CE models, those with binomial likelihoods, we show how to bring such models into the CE family, such that inferences in the modified model resemble the private variational Bayes algorithm as closely as possible, using the Pólya-Gamma data augmentation scheme. The iterative nature of variational Bayes presents a further challenge since iterations increase the amount of noise needed. We overcome this by combining: (1) an improved composition method for differential privacy, called the moments accountant, which provides a tight bound on the privacy cost of multiple VB iterations and thus significantly decreases the amount of additive noise; and (2) the privacy amplification effect of subsampling mini-batches from large-scale data in stochastic learning. We empirically demonstrate the effectiveness of our method in CE and non-CE models including latent Dirichlet allocation, Bayesian logistic regression, and sigmoid belief networks, evaluated on real-world datasets.

Word embeddings improve the performance of NLP systems by revealing the hidden structural relationships between words. Despite their success in many applications, word embeddings have seen very little use in computational social science NLP tasks, presumably due to their reliance on big data, and to a lack of interpretability. I propose a probabilistic model-based word embedding method which can recover interpretable embeddings, without big data. The key insight is to leverage mixed membership modeling, in which global representations are shared, but individual entities (i.e. dictionary words) are free to use these representations to uniquely differing degrees. I show how to train the model using a combination of state-of-the-art training techniques for word embeddings and topic models. The experimental results show an improvement in predictive language modeling of up to 63% in MRR over the skip-gram, and demonstrate that the representations are beneficial for supervised learning. I illustrate the interpretability of the models with computational social science case studies on State of the Union addresses and NIPS articles.

The iterative nature of the expectation maximization (EM) algorithm presents a challenge for privacy-preserving estimation, as each iteration increases the amount of noise needed. We propose a practical private EM algorithm that overcomes this challenge using two innovations: (1) a novel moment perturbation formulation for differentially private EM (DP-EM), and (2) the use of two recently developed composition methods to bound the privacy “cost” of multiple EM iterations: the moments accountant (MA) and zero-mean concentrated differential privacy (zCDP). Both MA and zCDP bound the moment generating function of the privacy loss random variable and achieve a refined tail bound, which effectively decrease the amount of additive noise. We present empirical results showing the benefits of our approach, as well as similar performance between these two composition methods in the DP-EM setting for Gaussian mixture models. Our approach can be readily extended to many iterative learning algorithms, opening up various exciting future directions.

We introduce a novel approach for estimating Latent Dirichlet Allocation (LDA) parameters from collapsed Gibbs samples (CGS), by leveraging the full conditional distributions over the latent variable assignments to efficiently average over multiple samples, for little more computational cost than drawing a single additional collapsed Gibbs sample. Our approach can be understood as adapting the soft clustering methodology of Collapsed Variational Bayes (CVB0) to CGS parameter estimation, in order to get the best of both techniques. Our estimators can straightforwardly be applied to the output of any existing implementation of CGS, including modern accelerated variants. We perform extensive empirical comparisons of our estimators with those of standard collapsed inference algorithms on real-world data for both unsupervised LDA and Prior-LDA, a supervised variant of LDA for multi-label classification. Our results show a consistent advantage of our approach over traditional CGS under all experimental conditions, and over CVB0 inference in the majority of conditions. More broadly, our results highlight the importance of averaging over multiple samples in LDA parameter estimation, and the use of efficient computational techniques to do so.

Arti Ramesh, Shachi H. Kumar, James Foulds, Lise Getoor. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

Dhanya Sridhar, James Foulds, Bert Huang, Lise Getoor, Marilyn Walker. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

Relational phrases (e.g., "got married to") and their hypernyms (e.g., "is a relative of") are central for many tasks including question answering, open information extraction, paraphrasing, and entailment detection.This has motivated the development of several linguistic resources (e.g.DIRT, PATTY, and WiseNet) which systematically collect and organize relational phrases.These resources have demonstrable practical benefits, but are each limited due to noise, sparsity, or size.We present a new general-purpose method, RELLY, for constructing a large hypernymy graph of relational phrases with high-quality subsumptions using collective probabilistic programming techniques.Our graph induction approach integrates small highprecision knowledge bases together with large automatically curated resources, and reasons collectively to combine these resources into a consistent graph.Using RELLY, we construct a high-coverage, high-precision hypernymy graph consisting of 20K relational phrases and 35K hypernymy links.Our evaluation indicates a hypernymy link precision of 78%, and demonstrates the value of this resource for a document-relevance ranking task.

Topic models have become increasingly prominent text-analytic machine learning tools for research in the social sciences and the humanities. In particular, custom topic models can be developed to answer specific research questions. The design of these models requires a non-trivial amount of effort and expertise, motivating general-purpose topic modeling frameworks. In this paper we introduce latent topic networks, a flexible class of richly structured topic models designed to facilitate applied research. Custom models can straightforwardly be developed in our framework with an intuitive first-order logical probabilistic programming language. Latent topic networks admit scalable training via a parallelizable EM algorithm which leverages ADMM in the M-step. We demonstrate the broad applicability of the models with case studies on modeling influence in citation networks, and U.S. Presidential State of the Union addresses.

Understanding the diffusion of information in social network and social media requires modeling the text diffusion process. In this work, we develop the HawkesTopic model (HTM) for analyzing text-based cascades, such as "retweeting a post" or "publishing a follow-up blog post". HTM combines Hawkes processes and topic modeling to simultaneously reason about the information diffusion pathways and the topics characterizing the observed textual information. We show how to jointly infer them with a mean-field variational inference algorithm and validate our approach on both synthetic and real-world data sets, including a news media dataset for modeling information diffusion, and an ArXiv publication dataset for modeling scientific influence. The results show that HTM is significantly more accurate than several baselines for both tasks.

Detecting unsolicited content and the spammers who create it is a long-standing challenge that affects all of us on a daily basis. The recent growth of richly-structured social networks has provided new challenges and opportunities in the spam detection landscape. Motivated by the Tagged.com social network, we develop methods to identify spammers in evolving multi-relational social networks. We model a social network as a time-stamped multi-relational graph where vertices represent users, and edges represent different activities between them. To identify spammer accounts, our approach makes use of structural features, sequence modelling, and collective reasoning. We leverage relational sequence information using k-gram features and probabilistic modelling with a mixture of Markov models. Furthermore, in order to perform collective reasoning and improve the predictive power of a noisy abuse reporting system, we develop a statistical relational model using hinge-loss Markov random fields (HL-MRFs), a class of probabilistic graphical models which are highly scalable. We use Graphlab Create and Probabilistic Soft Logic (PSL) to prototype and experimentally evaluate our solutions on internet-scale data from Tagged.com. Our experiments demonstrate the effectiveness of our approach, and show that models which incorporate the multi-relational nature of the social network significantly gain predictive performance over those that do not.

As the amount of recorded digital information increases, there is a growing need for flexible recommender systems which can incorporate richly structured data sources to improve recommendations. In this paper, we show how a recently introduced statistical relational learning framework can be used to develop a generic and extensible hybrid recommender system. Our hybrid approach, HyPER (HYbrid Probabilistic Extensible Recommender), incorporates and reasons over a wide range of information sources. Such sources include multiple user-user and item-item similarity measures, content, and social information. HyPER automatically learns to balance these different information signals when making predictions. We build our system using a powerful and intuitive probabilistic programming language called probabilistic soft logic, which enables efficient and accurate prediction by formulating our custom recommender systems with a scalable class of graphical models known as hinge-loss Markov random fields. We experimentally evaluate our approach on two popular recommendation datasets, showing that HyPER can effectively combine multiple information types for improved performance, and can significantly outperform existing state-of-the-art approaches.

When reviewing scientific literature, it would be useful to have automatic tools that identify the most influential scientific articles as well as how ideas propagate between articles.In this context, this paper introduces topical influence, a quantitative measure of the extent to which an article tends to spread its topics to the articles that cite it.Given the text of the articles and their citation graph, we show how to learn a probabilistic model to recover both the degree of topical influence of each article and the influence relationships between articles.Experimental results on corpora from two well-known computer science conferences are used to illustrate and validate the proposed approach.

There has been an explosion in the amount of digital text information available in recent years, leading to challenges of scale for traditional inference algorithms for topic models. Recent advances in stochastic variational inference algorithms for latent Dirichlet allocation (LDA) have made it feasible to learn topic models on very large-scale corpora, but these methods do not currently take full advantage of the collapsed representation of the model. We propose a stochastic algorithm for collapsed variational Bayesian inference for LDA, which is simpler and more efficient than the state of the art method. In experiments on large-scale text corpora, the algorithm was found to converge faster and often to a better solution than previous methods. Human-subject experiments also demonstrated that the method can learn coherent topics in seconds on small corpora, facilitating the use of topic models in interactive document analysis software.

Multi-instance (MI) learning is a variant of supervised learning where labeled examples consist of bags (i.e. multi-sets) of feature vectors instead of just a single feature vector. Under standard assumptions, MI learning can be understood as a type of semi-supervised learning (SSL). The difference between MI learning and SSL is that positive bag labels provide weak label information for the instances that they contain. MI learning tasks can be approximated as SSL tasks by disregarding this weak label information, allowing the direct application of existing SSL techniques. To give insight into this connection we first introduce multi-instance mixture models (MIMMs), an adaption of mixture model classifiers for multi-instance data. We show how to learn such models using an Expectation-Maximization algorithm in the case where the instance-level class distributions are members of an exponential family. The cost of the semi-supervised approximation to multi-instance learning is explored, both theoretically and empirically, by analyzing the properties of MIMMs relative to semi-supervised mixture models.