Abstract:

Probabilistic programming is a practice to model uncertainty by representing complex probabilistic models as simple programs. Various probabilistic programming systems automate the reasoning of these probabilistic programs with generic inference procedures. Benefiting from interpretability and efficiency, probabilistic programming has gained popularity in critical areas, like medical diagnosis, autonomous vehicles, and ecosystem management. However, models encoded as the probabilistic programs often rely on multiple assumptions of data. Violation of these assumptions may lead to the wrong result given by these programs. For models used in complex engineering systems for prediction or decision making, it's essential to assure the robustness. Therefore, we aim to automatically improve the robustness of probabilistic programs through automatic program transformation. We combine insights from machine learning and programming languages / program analysis. To assess the robustness of the models, we also design strategies to generate noisy data that may cause a mismatch. We evaluate the models with two inference algorithms: MCMC sampling and variational inference. Our evaluation shows that our automatic techniques lead to significant robustness improvement in different kinds of models.

Abstract:

Listen, Attend and Spell (LAS) maps a sequence of acoustic spectra directly to a sequence of graphemes, with no explicit internal representation of phones. This paper asks whether LAS can be used as a scientific tool, to discover the phone set of a language whose phone set may be controversial or unknown. Phonemes have a precise linguistic definition, but phones may be defined in any manner that is convenient for speech technology: we propose that a practical phone set is one that can be inferred from speech following certain procedures, but that is also highly predictive of the word sequence. We demonstrate that such a phone set can be inferred by clustering the hidden nodes activation vectors of an LAS model during training, thus encouraging the model to learn a hidden representation characterized by acoustically compact clusters that are nevertheless predictive of the word sequence. We further define a metric for the quality of a phone set (the sum of conditional entropy of the graphemes given the phone set and the phones given the acoustics), and demonstrate that according to this metric, the clustered- LAS phone set is comparable to the original TIMIT phone set. Specifically, the clustered-LAS phone set is closer to the acoustics; the original TIMIT phone set is closer to the text.

Abstract:

We consider clustering with the help of an oracle when all queries are made at once and the clusters are determined after all the responses are received. We determine the minimum number of queries required to completely cluster all items. We also consider active clustering with the help of an oracle when a number of queries are made in the first round and the remaining queries are made in the second round based upon the responses to the queries from the first round. We determine a lower bound for the number of queries required to completely cluster all items based upon an analysis of the number of queries made in the first round.

Abstract:

Finding similar looking galaxies can be hard to do by eye, and even harder when you have thousands of images to search through and compare. The Dark Energy Survey Interface allows users to find observational data on thousands of objects, and that framework could be helped by a service that finds visual matches. This API reduces the time and labor required to complete this task. Using this tool, astronomers studying galaxy formation and evolution can galaxies similar in shape and appearance to one they are studying to further explore relationships more efficiently than before. Additionally, inputting anomalous images could assist in finding objects where data is missing. This API is written in Python and uses tornado and Keras to match user inputted images of a galaxy to a database of galaxy images. It works by taking in an image and the number of return images wanted though an html form. Then the image is compressed into a vector using the Keras Autoencoder and compared to the vectorized images of many galaxies in a database made up of previously encoded image vectors. From there, the n closest match vectors are chosen using Euclidean distance and the corresponding images are templated into an html file that is then rendered. In the future, this API will be expanded to return more identifying information about each matching galaxy in addition to its image. IDs, coordinates, distance, and certain calculations would make the service even more useful to those comparing and contrasting similar galaxies and studying galaxy formation and evolution. The available dataset will also be expanded for more accurate results.

Abstract:

Imitation learning has been successfully used in many real-world scenarios such as autonomous driving, especially learning from video demonstrations. Unlike traditional imitation learning where both the observations and actions from an expert agent are provided to the student agent for training, imitation learning from observation (ILO) learns from observation only, which is more applicable as it is more intuitive and the actions are usually hard to detect accurately. However, one may encounter gaps in performance between the simulation results and the actual real-world results due to the discrepancy between training and test environments. In this paper, we propose a novel framework named Robust Imitation Learning from Observation (RILO) that aims to provide robustness for the student agent in an ILO setting. We introduce an adversary agent that competes with the student agent and maximally destabilizes the student agent by applying disturbances to the system. We jointly train the agent and the adversary simultaneously so that the student agent is reinforced and becomes more robust to the various conditions. We empirically demonstrate that RILO enhances the robustness of the system, in terms of its steady performance in test environments with different mass values from the training environment.

Abstract:

If you are smart, do you make good decisions? Most people believe the answer to this question is yes, but a growing body of literature suggests that intelligence and decision-making are different skills and just because you are intelligent does not necessarily imply you also make good decisions. To test this hypothesis, my research seeks to determine the brain regions which work in tandem and which work separately in these cognitive processes by examining both structural and functional connectivity of the brain. Decision-making was assessed through the Adult Decision Making Competence battery (A-DMC), which comprised six types of decision making. Both fluid and crystallized intelligence were assessed through three different tests. A graph theory metric, small world propensity, was used to characterize both functional and structural images of the brain. We implemented principal component analysis in R to identify factors of intelligence and decision making. We then correlated those factors with the small world propensity measure of structural and functional connectivity. Preliminary results provide evidence for an overlap between the brain regions involved in decision making and intelligence, but there is also evidence for brain regions which are distinct in these cognitive processes.

Abstract:

Text style transfer rephrases a text from a source style (e.g., informal) to a target style (e.g., formal) while keeping its original meaning. Despite the success existing works have achieved using a parallel corpus for the two styles, transferring text style has proven significantly more challenging when there is no parallel training corpus. In this paper, we address this challenge by using a reinforcement-learning-based generator-evaluator architecture. Our generator employs an attention-based encoder-decoder to transfer a sentence from the source style to the target style. Our evaluator is an adversarially trained style discriminator with semantic and syntactic constraints that score the generated sentence for style, meaning preservation, and fluency. Experimental results on two different style transfer tasks (sentiment transfer and formality transfer) show that our model outperforms state-of-the-art approaches. Furthermore, we perform a manual evaluation that demonstrates the effectiveness of the proposed method using subjective metrics of generated text quality.

Abstract:

Nearly every course at The University of Illinois utilizes office hours for students to receive supplemental instruction (SI) outside of lecture. For many large courses at Illinois, an online self-service queueing system is used to facilitate the logistics of holding large office hours. When a student arrives at office hours facilitated by the queue, she adds herself to the online queue and is helped by the next available instructor or teaching assistant. This work in progress combines the log files generated by the online queueing system with student scores on graded assessments in order to determine how going to office hours impacts a student’s grade. The log files contain data such as student enqueue time, student dequeue time, when the student started getting help, and their question topic. The grade data contains grades for each individual assignment in the course, as well as the student’s final grade. Our initial work focuses on the data from the Data Structures course across one academic year (n=1,238 students). We believe this work is the largest observational analysis of the impact of office hours on graded assessments. Our findings suggest that there are new relationships that need further examination. The first major relationship we found was that office hour attendance provides an increase in a student’s score on upcoming graded homework, but it does not help increase a student's score on the exams. The second relationship is that the later a student goes to office hours during an assignment’s period, the lower the student’s grade on that assignment. With this research, we hope to improve the learning experience, which will help improve students’ exam scores and problem-solving abilities.

Abstract:

The term "tidy data" first entered the Data Science vernacular with Hadley Wickham's 2014 paper "Tidy Data" and accompanying R packages that would aid users in the data tidying process. However, the underlying theory behind "tidy data" is Codd's relational model, which itself is based on first order predicate logic and the theory of relations. As Wickham states in his paper, the principles of Tidy Data are those of Codd's third normal form, reframed for statisticians. The original 3 rules of tidy data (2014) were: "each variable forms a column," "each observation forms a row," and "each type of observational unit forms a table". Wickham later revised the third rule (in 2016) to be "each value must have its own cell". Similarly, the rules for a database in third normal form can be summarized as: all attributes are atomic, every non-prime attribute is functionally dependent on the (entire) primary key, and no non-prime attribute is functionally dependent on another non-prime attribute (no transitive dependencies). The tidy data philosophy is the foundation for the collection of R packages called the "tidyverse", which allow users to easily transform, model, and visualize data. But while the tidyverse dominates the R data science ecosystem, there is not a true equivalent in the Python data science ecosystem. For python users, most data wrangling tasks are accomplished with the pandas library. Data tidying is just one of the many tasks that could be described as data wrangling... but in my opinion it is the most important. The pandas library is extremely flexible and can manipulate tabular datasets into whatever form a user needs. However, this flexibility also means that it may be difficult for non-experts to fully utilize the library. This poster will demonstrate recipes that utilize the pandas library to tidy (normalize) datasets. The recipes will allow users to explore their datasets in order to discover the pre-existing functional dependencies between attributes (columns), primary (and candidate) keys, and multivalued attributes; transform their datasets by decomposing the table into new tables with proper functional dependencies, creating new identifiers, separating multivalued attributes, and otherwise normalizing the dataset(s); and finally verify that the new set of tables obey all necessary uniqueness constraints (primary keys), integrity constraints (foreign keys), check constraints, and otherwise conform to a relational model such that they could be loaded into a relational database. The recipes will be demonstrated on two data sources from my domain of political science which desperately need to be tidied: the Correlates of War project datasets and the UCDP/PRIO Armed Conflict dataset.

Abstract:

Social and organizational networks are often hierarchical by nature, where individuals belong to multiple groups and some groups may be harder to reach than others. For example, in organizational contexts, individuals may belong to different occupational roles with varying levels of authority and status that influence their network positions. Extant literature in organizational communication and status theory suggest connections between communication dynamics and formal hierarchy, in which conversations between authority and subordinates are distinctive from conversations between individuals of the same role. Stemming from prior theories, we posit that communication dynamics can be used to infer an individual's hierarchical position in the network, and we use sentiment as an indicator of communication among individuals. Specifically, we propose a sentiment-based ranking algorithm that utilizes PageRank in a signed directed graph. We compute sentiment scores for all actor pairs in a real-world organizational network (i.e. Enron email communication) which originally consists of 491 nodes and 7,344 edges. Preliminary triadic analysis reveals that sentiment in communication is dominantly positive, and regulated by transitive relations (e.g., A sends an email with positive sentiment towards B, but B sends an email with negative sentiment towards C, and to be consistent with B, A also sends an email with negative sentiment towards C). We calculated PageRank score for each individual from the incoming edges they received and the sentiment scores directed at them with introducing learning rate. We hypothesize that individuals at higher levels of the hierarchy (specifically board of directors, executive management, senior management at Enron) will have higher PageRank scores than individuals at lower levels (traders, specialists, associates at Enron). We further distinguish communities that potentially belong to different levels of hierarchy.

Abstract:

Observed social networks are often considered as proxies for underlying social networks. The analysis of observed networks oftentimes involves the identification of influential nodes via various centrality metrics. Our work is motivated by recent research on the investigation and design of adversarial attacks on machine learning systems. We apply the concept of adversarial attacks to social networks by studying how an adversary can minimally perturb the observed network structure. The goal of the adversary is to modify the ranking of nodes according to centrality measures. This can represent the attempts of an adversary to boost or demote the degree to which others perceive them as influential or powerful. It also allows us to study the impact of adversarial attacks on targets and victims, and to design metrics and security measures that help to identify and mitigate adversarial network attacks. We conduct a series of experiments on synthetic network data to identify attacks that allow the adversarial node to achieve their objective with a single move. We test this approach on different common network topologies and for common centrality metrics. We find that there is a small set of moves that result in the adversary achieving their objective, and this set is smaller for decreasing centrality metrics than for increasing them. These results can help with assessing the robustness of centrality measures. The notion of changing social network data to yield adversarial outcomes has practical implications, e.g., for information diffusion on social media, influence and power dynamics in social systems, and improving network security. Attacks on different networks have different implications: for example, in a flight transportation network, demoting an airport betweenness centrality means that an airport loses its place as a major transfer center. In social media networks, increasing an account's closeness centrality will help speed up information propagation (posts) to other users of the social media. These real-world implications can provide insights to improvements in network stability and security.

Abstract:

Although breakthroughs in deep learning technologies have seen great advancements in practical applications of computer vision recognition tasks, for example, Image classification, problems of low classification accuracy and weak generalizability of models still exist. In addition, while significant developments and successes have been made in solving these classification problems using a class of deep learning networks called Deep Convolutional Neural Networks (DCNN), more opportunities still exist in examining other newer architectures. Here, we propose to use ideas from prior large networks in order to leverage their successes and behavioral performances. To this end, we developed a dense-inception model to investigate salient problems such as model performance, learned representations etc. We test our novel model on a large medical image dataset, specifically from a large dataset of Chest X-rays called ChestXpert. Further, the model was trained from scratch and will be experimented using transfer learning (TL) methods such as fine-tuning techniques. The technical implementation includes the use of mini-batch, stochastic gradient descent, dropout and L2 regularization, batch normalization, optimization of epochs, and data augmentation techniques. We use a GPU environment with a TITAN V 12GB with two cards for the training and modeling process. The model is implemented using Keras library utilizing the TensorFlow backend. The primary objective is to produce innovations that maximize the use of multi-scale feature learning capable of handling large datasets, for example, the Chest X-ray images. Next, we plan to incorporate domain adaption using auto-encoders to learn deep features leveraging the properties of DCNNs and extend to classify other complex medical images such as CT and MRI scans hence maximizing on transferable features. Additionally, future work involves using other modern deep learning networks such as generative adversarial networks (GANs) to further investigate behavioral performances on medical image classification problems.

Abstract:

Text networks are promising in scientific domains, such as rumor detection, social network analysis, and scientific scholar retrieval. Researchers can apply novel algorithms on such networks to uncover connections of entities and make predictions. However, to construct reliable large text networks is challenging in practice. The performance of entity detection, relation extraction, and the selection of network construction methods all affect the final text networks. Our overall goal is to identify empirically-validated best practices in constructing relevant text networks using widely-used natural language processing (NLP) tools. We employ a two-step process in relation extraction: (1) identifying nodes (entities) and (2) determining relationships between nodes. In this poster, we present preliminary results on the first step. To determine the best practice for identifying nodes, we tested three widely-used NLP libraries with named entity recognition implementations: spaCy, NLTK, and Stanford NER. Results on the ACE Newswire 2005 dataset show that NLTK and spaCy have similar performance (24% and 23% recall, respectively), while the Stanford NER only obtained 7% recall. NLTK performs best on recognizing geopolitical entities (49% recall) and persons (15%), while spaCy performs best on organizations (30%), locations (9%), and facilities (2%).