Search

This paper introduces Prio+, a privacy-preserving system for the collection of aggregate statistics, with the same model and goals in mind as the original and highly influential Prio paper by Henry Corrigan-Gibbs and Dan Boneh (NSDI 2017). As in the original Prio, each client holds a private data value (e.g. number of visits to a particular website) and a small set of servers privately compute statistical functions over the set of client values (e.g. the average number of visits). To achieve security against faulty or malicious clients, unlike Prio, Prio+ clients use Boolean secret-sharing instead of zero-knowledge proofs to convince servers that their data is of the correct form and Prio+ servers execute a share conversion protocol as needed in order to properly compute over client data. This allows us to ensure that clients' data is properly formatted essentially for free, and the work shifts to novel share-conversion protocols between servers, where some care is needed to make it efficient. Our overall approach is simpler than Prio and our Prio+ strategy reduces the client's computational burden by at least two orders of magnitude (or more depending on the statistic) while keeping server costs comparable to Prio. Prio+ permits computation of exactly the same wide range of complex statistics as the original Prio protocol, including high-dimensional linear regression over private values held by clients. We report detailed benchmarks of our Prio+ implementation and compare these to both the original Go implementation of Prio and the Mozilla implementation of Prio. Our Prio+ software is open-source and released with the same license as Prio.

This paper describes the process of development and evaluation of an open educational resource (OER) e-module on the Central Limit Theorem written for an Introductory Statistics college-level course. The purpose of this project is two-fold. First, the e-module bridges the knowledge gap between introductory topics and Hypothesis Testing – one of the most challenging concepts in Statistics. Second, the project focuses on developing tools that allow instructors to analyze the effectiveness of the module and reveal student patterns of interaction with the platform. The overall goal of the project is to improve the quality of open educational resources, provide students/instructors with additional study materials in response to rising cost for textbooks and higher education, and provide more data for further research on student behavior while interacting with e-textbooks. The interactive e-module was developed using LaTeX markup language and Overleaf editor, uploaded to the XIMERA platform and tested on two sections of MTH 140, a college-level Statistics course. Once the experiment has been performed and the data collected, the results were analyzed using Python programming language. As a result of the study, some tools for analysis of user data have been developed, and an OER has been created.

Melanoma is an incredibly dangerous form of skin cancer. In 2023 alone, the American Cancer Society estimates that about 7,990 Americans will die from new melanoma cases. Image analysis tools that automate the diagnosis of melanoma will improve dermatologists' diagnostic accuracy and should provide a significantly better detection of melanoma. In this paper, we present an application of modern deep learning architectures using the images of skin lesions in combination with patient-level contextual information to detect the presence of melanoma. More specifically, we used the latest variant of the residual network, the ResNeSt model, to leverage its deep architecture and channel-wise attention mechanisms for melanoma detection. In parallel, we used a multi-layer perceptron model to process patient-level data, and fused its output with the extracted features from the ResNeSt model. To compare and contrast performances, we trained a similar model using a smaller, more generic, convolutional neural network in addition to the larger ResNeSt network. After properly tuning all associated hyperparameters, our primary model obtained a balanced accuracy of 0.853 and a ROC-AUC score of 0.93. The split-attention network was shown to improve the total ROC-AUC by approximately 9%, suggesting that our model was able to successfully leverage the ResNeSt design to extract the most important features and most relevant cross-feature interactions for the detection of melanoma.

In this paper we conduct transfer learning on the Faster R-CNN ResNet-50 FPN V2 object detection model from Benchmarking Detection Transfer Learning with Vision Transformers [12] to a new problem space: detecting what parts of a webpage can be clicked on. The model predicts the bounding boxes and class labels for what part of a website image can be clicked on, and what the result will be. We found that training the feature extraction layers of the model was detrimental to the learning (354% higher loss after 10 epochs), but that training the RPN layers was not detrimental. We additionally found that the most effective resizing for website imagery was to 224 x 224 pixels. Since the meta trainer we developed did not produce usable data, we experiment on using the data diversity approach in active learning to select the most representative images to train on. Compared to our control, the model trained on the active learning dataset scored 12.5% better on the validation data. However, with such a small dataset of 10 labeled images, the model was unable to generalize and did not predict any of the 15 objects on a labeled image outside the training set.

Data scientists work with large-scale sensitive data, which inevitably leads to privacy risks. Differential Privacy (DP) is a mathematical definition of privacy that aims to mitigate privacy risks inherent in data analysis and machine learning. OpenDP, an open-source software DP tool, allows the government, industry, and academic institutions to share sensitive data to researchers or the public while preserving privacy. An active research question in DP literature is how to bound the total privacy loss of a sequence of DP computations. Most known DP theorems and the OpenDP library require the privacy parameters of each computation to be fixed in advance. However, this prevents the design of privacy preserving machine learning algorithms that change the privacy budget on the fly. For the adaptive parameter setting, privacy filters and odometers are two objects designed to model the total privacy loss. In this paper, we extend the programming framework for OpenDP library to handle DP composition under adaptive privacy budgets through Renyi Differential Privacy (RDP). To do so, we construct Renyi filter and odometer and prove its privacy guarantees by generalizing RDP Adaptive Composition. To generalize our Odometer results, we implement a constructor that converts any Odometer to a Filter. Our results allow for the real world DP deployment of ML algorithms and interactive query interfaces that adaptively update the privacy budget.

Reliably diagnosing and evaluating generative models remains an open problem. The most common metric employed to measure the performance of image generating models has been the Fréchet Inception Distance (FID) score. It has been shown to correlate well with human judgement and it is sensitive to mode collapse. However, recent papers have demonstrated issues with the metric, such as poor performance on certain tasks, long computation time, and failure to match human perception. We propose a novel formulation, Class-Aware Latent Distance (CALD), to alleviate these issues. The proposed score can be applied across domains, is quicker to compute, and matches human judgement where FID fails. To demonstrate the capability of this approach, we perform an empirical study on three image generation tasks and compare our metric to FID and Intrinsic Multiscale Distance (IMD). Across these experiments, we show that the proposed metric offers significant speedup over existing methods, is more reliable, and can be used to diagnose and perform early stopping during the training of Generative Adversarial Networks.

The purpose of this paper is to analyze how to predict a business’s operation status during the spread of COVID-19 between the years 2020 to 2022. Exploratory Data Analysis techniques including barplot, geographical visualization, and word cloud are utilized. Three models are chosen to fit and predict business operation status: Logistic Regression, Support Vector Machine, and Bidirectional Long Short Term Memory. After tuning the probability cutoff threshold to 0.0510 with the F1 score, the logistic regression using the solver Newton's Conjugate Gradient generates an accuracy score of around 0.8120. Support Vector Machine algorithm leads to an accuracy score of around 0.9763. By using GloVe for word embedding, bidirectional LSTM results in a loss of 0.1113 and an accuracy score of 0.9758 after 2 epochs. From the performance of the three models, SVM seems to be the most effective and efficient in predicting business operation status through user reviews.

This paper proposes a conditional version of Divergence Triangle [1] as a framework to train generator, energy-based and inference models jointly with the information of labels, where the learning of the above three models are integrated perfectly in a unified probabilistic formulation. Experiments demonstrate that, within this one framework, we are able to complete the following tasks together, (1) control the fine-grained categories to generate realistic images, (2) obtain the meaningful representation of observed data in the low dimensions, and also (3) conduct label classification on unobserved data. Additionally, I also discuss a possible extension on Conditional Divergence Triangle model at the end of this paper for future work.

Cumulative probability models have been introduced as a flexible alternative to linear models. However, they are computationally heavy especially when the number of distinct outcomes is large. We introduced and implemented three methods to address this problem. In the divide-and-combine method, the data are partitioned into subsets and the models are fit in parallel; in the binning and rounding methods, the outcomes are either grouped into bins or rounded (to certain significant digits or decimal places) before the models are fit. We implemented these approaches, built simulation programs, and developed an R package. We describe various challenges during the implementation and present their solutions. Examples include using ExitStack for data partition, and implementing an algorithm to merge all step functions to obtain the final estimates. We also briefly describe the results from our paper and an application to a large dataset.

This dissertation consists of several papers. First, we start off introducing domain adaptation theory and briefly introduce optimal transport. Such an introduction allows the reader to understand why studying problems in optimal transport theory is so valuable. Our first key result establishes bounds between regularized and unregularized optimal transport. Instead of using an entropic regularization, which is used in the Sinkhorn divergence, we regularize using dual potentials in a reproducing kernel Hilbert space. After this, we derive sample complexity bounds for the regularized optimal transport problem, and we show this is a substantial improvement over unregularized optimal transport. With these two results, one can approximate the theoretical optimal transport distance. Next, we prove the first and second moments of the source and target distributions are enough to determine explicitly the optimal transport map and also that this is a linear mapping. Furthermore, we propose an alternative regularization for the transport map between two distributions. After this, we briefly diverge from optimal transport theory and introduce work on prior elicitation. In particular, we extend a result from \cite{Pinelis2017} on non-asymptotic bounds for maximum likelihood estimators to that for M-estimators. Crucially, we show sufficient assumptions for these to hold and use these to theoretically justify our prior elicitation objective. Last, we return to optimal transport and introduce a variant to compare multiple probability measures, which we call sliced multi-marginal optimal transport. There, we propose a paradigm based on random one-dimensional projections.

Modern data science calls for statistical inference algorithms that are both data-efficient and computation-efficient. We design and analyze methods that 1) outperform existing algorithms in the high-dimensional setting; 2) are as simple as possible but efficiently computable. One line of our work aims to bridge different fields and offer unified solutions to multiple questions. Consider the following canonical statistical inference tasks: distribution estimation, functional estimation, and property testing, sharing the model that provides sample access to an unknown discrete distribution. In a recent paper in NeurIPS '19, we showed that a single, simple, and unified estimator – profile maximum likelihood (PML), and its near-linear time computable variant are sample-optimal for estimating multiple distribution attributes. The result covers 1) any appropriately Lipschitz and additively separable functionals; 2) sorted distribution; 3) R\'enyi entropy; 4) $\ell_2$ distance to the uniform distribution, yielding an optimal tester for distributions' closeness. This work makes PML the first unified sample- and time-optimal method for the learning tasks mentioned above. A single algorithm with such broad applicability is \emph{universal}. Another line of our work focuses on instance-optimal learning that designs algorithms with near-optimal guarantees for every possible data input. A flagship problem is distribution estimation over discrete or continuous domains, where ordering and geometry play an essential role. Going beyond worst-case guarantees, researchers designed algorithms that compete with a genie estimator that knows the actual distribution but is reasonably restricted. To obtain state-of-the-art algorithms for both tasks, we leveraged the simple but nontrivial idea of ``data binning''. For discrete settings, we group symbols that appear the same number of times. And for continuous settings, we partition the real domain and separate symbols according to pre-designed local quantiles. The respective algorithms run in near-linear-time, achieve the best-known estimation guarantees regarding the genie estimators, and appear in ICML'19 and NeurIPS '20. A genie-like algorithm adaptive to almost every data sample is \emph{competitive}.We present a comprehensive understanding of universal and competitive algorithms for multiple fundamental learning problems. Our ideas and techniques may shed light on key challenges in modern data science and numerous applications beyond the scope of this dissertation.

The problem of embedding very large information networks into low-dimensional vector spaces is useful in many tasks such as visualization, node classification, and link prediction. This paper studies the novel network embedding method called the ''LINE'' , which optimizes a carefully designed objective function that preserves both the local and global network structures. In order to rule out the instability on border vertices' embeddings and its influence on core vertices, we only compute the core graph with LINE after peeling out those border vertices. Then we compute the embedding of peeled nodes with locally updated process. This paper also tried to interpret and visualize the local update process with logistic regression, and optimize the local update process by adding prior and intercept to the objective function. Finally, this paper demonstrate the embeddings on several multi-label network classification tasks for social networks such as BlogCatalog and YouTube. Our results show that the optimized LINE outperforms the initial methods 5% in F1-score with YouTube dataset and speed up the convergence time.

This thesis focuses on the role investor type and sentiment play in financial markets, using data from social media. First paper investigates the effect of the interaction between asset maturity and liquidity restrictions in ``on-the-run" phenomenon, using asset markets with search frictions. Under the presence of search frictions, investors would not prefer holding assets with very short time to maturity, since they will need to go back to the market and search for a counterpart to buy new assets every time their assets mature, incurring a search cost. However, they also would not want to hold assets with very long time to maturity due to liquidity considerations. An asset search model is set up to determine asset choices of investors with different liquidity preferences. Model considers two assets that differ in their maturities and two investor types who differ in their liquidity preferences. Main finding of this paper is that liquidity cost matters in the presence of search frictions as the model predicts a separating equilibrium where high type agents choose the long term asset and the low type agents choose the short term asset. When the two assets have the same time-to-maturities, the same separating equilibrium is obtained. Spread of the long term asset is found to be higher than that of the short term asset, which goes in line with the data and hence this paper shows that ``on-the-run" phenomenon can be explained by higher search frictions in the off-the-run markets and investors with different liquidity preferences.Second paper predicts intra-day foreign exchange rates by making use of trending topics from Twitter, using a sentiment based topic clustering algorithm. Twitter trending topics data provide a good source of high frequency information, which would improve the short-term or intra-day exchange rate predictions. This project uses an online dataset, where trending topics in the world are fetched from Twitter every ten minutes since July 2013. First, using a sentiment lexicon, the trending topics are assigned a sentiment (negative, positive, or uncertain), and then using a continuous Dirichlet process mixture model, the trending topics are clustered regardless of whether they are explicitly related to the currency under consideration. This unique approach enables to capture the general sentiment among users, which implicitly affects the currencies. Finally, the exchange rates are estimated using a linear model which includes the topic based sentiment series and the lagged values of the currencies, and a VAR model on the topic based sentiment time series. The main variables of interest are Euro/USD, GBP/USD, Swiss Franc/USD and Japanese Yen/USD exchange rates. The linear model with the sentiments from the topics and the lagged values of the currencies is found to perform better than the benchmark AR(1) model. Incorporating sentiments from tweets also resulted in a better prediction of currency values after unexpected events.Third paper investigates the behavior of Reddit's news subreddit users and the relationship between their sentiment on exchange rates. Using graphical models and natural language processing, hidden online communities among Reddit users are discovered. The data used in this project are a mixture of text and categorical data from a news website. It includes the titles of the news pages, as well as a few user characteristics, in addition to users' comments. This dataset is an excellent resource to study user reaction to news since their comments are directly linked to the webpage contents. The model considered in this paper is a hierarhical mixture model which is a generative model that detects overlapping networks using the sentiment from the user generated content. The advantage of this model is that the communities (or groups) are assumed to follow a Chinese restaurant process, and therefore it can automatically detect and cluster the communities. The hidden variables and the hyperparameters for this model can be obtained using Gibbs sampling.

Principal Component Analysis has been extensively used in the computer vision field as a method of capturing orthogonal axes of large variability in high-dimensional data sets. Computer vision scientists have come up with reconstructive models which capture the most distinguished features of a human face using Principal Component Analysis, known as “Eigenfaces”. Several papers have approached the problem of facial recognition using standard PCA, however very few provide a detailed comparison on the different non-linear kernels which can be used in place of the traditional linear approach. The aim of this paper is to introduce several non-linear kernel functions to the human recognition problem, by working with a set of radial basis kernels, a logarithmic kernel, a Cauchy kernel, and a polynomial kernel. We perform a model assessment for each kernel using a parameter tuning method which minimizes reconstruction error, and display reconstruction plots for each kernel method. We also capture influential physical features of the images in the high-dimensional space (the Eigenface) for each kernel and compare reconstructed and original images, by capturing the Frebenius (L2) norm between test and original image data.

Biological processes are often far too complex to predict. For those phenomenon that still evade understanding, it is helpful to visualize a black box—a system with clear inputs and outputs, but an unknowable, labyrinthine set of interior mechanisms to produce this transformation. To determine what happened, controlled experiments are often run on a culture of cells or tissues using a multi-well plate format. This structure is relatively simple and can be applied in most laboratory environments. This paper discusses newly built technologies that take this definition of an experiment and adapt it to detect genotypes as well as phenotypes; promisingly, these technologies are highly multiplexible by using a modified version of Fluoresecnt in situ Sequencing (FISSEQ).

This paper uses player tracking data from a subset of games from the 2015-16 NBA season to estimate the relationship between energy expenditure and player performance. Using locational tracking data, several measures of energy expenditure are calculated, along with one baseline measure. To account for each individual player's skill level, a ridge regression is performed and the residuals are used as the dependent variable in a second regression. Binning is used to estimate the relationship between measures of energy expenditure and player performance, and the coefficients from binning are subsequently used to estimate parametric transformations of energy expenditure that are included in the final model. Finally, generalized additive models (GAMs) are used to estimate the relationship between energy expenditure and performance for individual players. I find three notable results. First, energy expenditure is a significant predictor of a player's deviation from their average skill level. Second, player tracking data can be used to generate measures of energy expenditure that allow for significantly more accurate predictions of player performance than measures not derived from player tracking data. Third, there is strong evidence that players require a period of time when they first enter the game to warm up, during which they perform at a lower level than normal.

In recent years, Social Network Service (SNS) is a novel, popular way to make friends andconvey information online. Therefore, the analysis of network data has attracted a lot ofattention. It is an area that is rapidly growing, both with Statistics and Computer Science.This paper rst provides a summary of statistical methods used in network data analysis,including basic denitions, measurements, and descriptive statistics. We then introduce theExponential Random Graph Model to t network data. Secondly, we dig into a more specicarea of network analysis: Community Detection. We discuss two dierent methods to explorethe community stucture, one is Louvain algorithm and the other is Mixed MembershipStochastic Blockmodels. After that, we combine the community identication with a two-stage user similarity algorithm to build a friend recommendation method. In the empiricalstudy section, we apply this method to a real-world dataset and evaluate its performacethrough specic measurements.

Attention mechanisms have recently been introduced in deep learning for various tasks in natural language processing and computer vision. But despite their popularity, the ``correctness'' of the implicitly-learned attention maps has only been assessed qualitatively by visualization of several examples. In this paper we focus on evaluating and improving the correctness of attention in neural image captioning models. Specifically, we propose a quantitative evaluation metric for how well the attention maps align with human judgment, using recently released datasets with alignment between regions in images and entities in captions. We then propose novel models with different levels of explicit supervision for learning attention maps during training. The supervision can be strong when alignment between regions and caption entities are available, or weak when only object segments and categories are provided. We show on the popular Flickr30k and COCO datasets that introducing supervision of attention maps during training solidly improves both attention correctness and caption quality.

In this paper, we propose an approach to classify action sequences. We observe that in action sequences the critical features for discriminating between actions occur only within sub-regions of the image. Hence deep network approaches will address the entire image are at a disadvantage. This motivates our strategy which uses static and spatio-temporal visual cues to isolate static and spatio-temporal regions of interest (ROIs). We then use weakly supervised learning to train deep network classifiers using the ROIs as input. More specifically, we combine multiple instance learning (MIL) with convolutional neural networks (CNNs) to select discriminative action cues. This yields classifiers for static images, using the static ROIs, as well as classifiers for short image sequences (16 frames), using spatio-temporal ROIs. Extensive experiments performed on the UCF101 and HMDB51 benchmarks show that both these types of classifiers perform well individually and achieve state of the art performance when combined together. We also show qualitatively that our ROIs (selected by the algorithms) capture the most relevant parts of the image sequences.

Many problems require experimental means to determine the best solutions. Experimenters must use statistics to analyze the results of the experiments. Proper statistical evaluation of the results could be easier, with some assistance. This paper proposes Computer Assisted Treatment Evaluation, a program to assist people in the statistical evaluation of the results of experiments in the area of computer algorithm analysis. This provides a framework upon which even more extensive evaluation can be built. An experiment evaluating several sorting algorithms provides an example.