All Seminars
| Title: Cell Type Identification in Single-cell Genomics and its Applications |
|---|
| Defense: Computer Science |
| Speaker: Wenjing Ma, Emory University |
| Contact: Hao Wu, Hao.wu@emory.edu |
| Date: 2023-04-27 at 3:00PM |
| Venue: MSC E406 |
 Download Flyer Add to Calendar |
| Abstract: Advances in techniques for measuring genomics in cell-level resolution provide great opportunities to reveal the cell heterogeneity. When dealing with single-cell genomics sequencing data, the most fundamental and critical step is to accurately identify cell types (celltyping). Once cell types are identified, the understanding of cell signatures, cellular composition and cell dynamics can broaden the knowledge of biological processes and potentially cure diseases. Traditional approaches of celltyping in single-cell genomics data are based on unsupervised clustering, expert’s knowledge, and manual curation, which are labor intensive. With continuous accumulation of high-quality single-cell genomics data, supervised celltyping becomes more and more popular due to its accuracy, robustness, and efficiency. In this seminar, the speaker will first introduce a benchmark study in supervised celltyping in single-cell RNA-sequencing data. Based on the experience gained from the benchmark study, the speaker will then introduce the supervised celltyping methods specifically developed for single-cell chromatin accessibility (scATAC-seq) data. Finally, with the accurately identified cell types, the speaker will introduce a method developed to identify cellular activity in bulk differential expression study with the usage of cell-type-specific marker information. |
| Title: Defensive Machine Learning Techniques for Countering Adversarial Attacks |
|---|
| Defense: Computer Science |
| Speaker: Fereshteh Razmi, Emory University |
| Contact: Li Xiong, lxiong@emory.edu |
| Date: 2023-04-20 at 2:00PM |
| Venue: https://zoom.us/my/lxiong |
| Download Flyer Add to Calendar |
| Abstract: The increasing reliance on machine learning algorithms has made them a target for exploiting vulnerabilities in these systems and launching adversarial attacks. The attacker in these attacks manipulates either the training data or test data, or both, known as a poisoning attack, adversarial example, or backdoor attack, respectively. They primarily aim to disrupt the model's classification task. In cases where the model is interpretable, the attacker may target the interpretation of the model's output. These attacks can have significant negative impacts; therefore, it is crucial to develop effective defense methods to protect against them. Current defense methods have limitations. Outlier detectors, used to identify and mitigate poisoning attacks, require prior knowledge of the attack and clean data to train the detector. Robust defense methods show promising results in mitigating backdoor attacks, but their effectiveness comes at the cost of decreased model utility. Furthermore, few defense methods have addressed adversarial examples that target the interpretation of the model's output. To address these limitations, we propose defense methods that protect machine learning models from adversarial attacks. Our methods include an autoencoder-based detection approach to identify various untargeted poisoning attacks. We also provide a comprehensive comparative study of differential privacy approaches and suggest new approaches based on label differential privacy to defend against backdoor attacks. Lastly, we propose a novel attack and defense method to protect the interpretation of a healthcare-related machine learning model. These approaches represent significant progress in the field of machine learning security and have the potential to protect against a wide range of adversarial attacks." |
| Title: Contextual Embedding Representation for Dialogue Systems |
|---|
| Defense: Computer Science |
| Speaker: Zihao Wang, Emory University |
| Contact: Jinho Choi, jinho.choi@emory.edu |
| Date: 2023-03-30 at 11:00AM |
| Venue: MSC E306 |
| Download Flyer Add to Calendar |
| Abstract: Context is a crucial element for conversational agents to conduct natural and engaging conversations with human users. By being aware of the context, a conversational agent can capture, understand, and utilize relevant information, such as named entity mentions, topics of interest, user intents, and emotional semantics. However, incorporating contextual information into dialogue systems is a challenging task due to the various forms it can take, the need to decide which information is most relevant, and how to organize and integrate it. To address these challenges, this thesis proposes exploring and experimenting with different contextual information in the embedding space across different models and tasks. Furthermore, the thesis develops models that overcome the limitations of state-of-the-art language models in terms of the maximum number of tokens they can encode and their incapacity to fuse arbitrary forms of contextual information. Additionally, diarization methods are explored to resolve speaker ID errors in the transcriptions, which is crucial for training dialogue data. The proposed models address the challenges of context integration into retrieval-based and generation-based dialogue systems. In retrieval-based systems, a response is selected and returned by ranking all responses from different components. A contextualized conversational ranking model is proposed and evaluated on the MSDialog benchmark conversational corpus, where three types of contextual information are leveraged and incorporated into the ranking model: previous conversation utterances from both speakers, semantically similar response candidates, and domain information associated with each candidate response. The performance of the contextual response ranking model exceeded state-of-the-art models in previous research, showing the potential to incorporate various forms of context into modeling. In generation-based systems, a generative model generates a response to be returned to the conversing party. A generative model is built on top of the Blenderbot model, overcoming its limitations to integrate two types of contextual information: previous conversation utterances from both conversing parties and heuristically identified stacked questions that tackle repetition and provide topical diversity in dialogue generations. The models are trained on an interview dataset and evaluated on an annotated test set by professional interviewers and students in real conversations. The average satisfaction score from professional interviewers and students is 3.5 out of 5, showing promising future applications. Additionally, to better understand topics of interest, topical clustering and diversity are investigated by grouping topics and analyzing the topic flow in the interview conversations. Frequent occurrences of some clusters of topics give a clear presentation of what scopes of topics an interview would touch on while maintaining a great selection of unique topics for individuals. Based on this observation, another generative model architecture integrating topical information is proposed that generates the next topic of interest in the conversation flow in parallel to generating utterances. This work is ongoing, with the expectation of improving the performance of the previous generative model. Day/time: March 30th, 11:00 am - 12:30 pm Room: Math CS E306 Zoom Option: https://us02web.zoom.us/j/9910064905?pwd=aThaYVd1eFBkRWpwQ2xibFFneHIzUT09 Meeting ID: 991 006 4905 Passcode: 290751 |
| Title: Computational Structures as Neural Symbolic Representation |
|---|
| Defense: Computer Science |
| Speaker: Han He, Emory University |
| Contact: Jinho Choi, jinho.choi@emory.edu |
| Date: 2023-03-30 at 5:00PM |
| Venue: Emerson E101 |
| Download Flyer Add to Calendar |
| Abstract: Although end-to-end neural models have been dominating Natural Language Processing for both performance and flexibility, critics have recently drawn attention to their poor generalization and lack of interpretability. Conversely, symbolic paradigms such as Abstract Meaning Representation (AMR) are humanly comprehensible but less flexible. In response, we propose Executable Abstract Meaning Representation (EAMR) as a reconciliation of both paradigms. EAMR is a neural symbolic framework that frames a task as a program, which interactively gets generated, revised and executed. In our novel definition, execution is a sequence of transforms on AMR graphs. Through a hybrid runtime, EAMR learns the automatic execution of AMR graphs, yet it also allows for the integration of handcrafted heuristics, knowledge bases and APIs. EAMR can be used in many applications such as dialogue understanding and response generation. https://emory.zoom.us/j/91398763361 |
| Title: Attention-enhanced Deep Learning Models for Data Cleaning and Integration |
|---|
| Defense: Computer Science |
| Speaker: Jing Zhang, Emory University |
| Contact: Joyce Ho, joyce.c.ho@emory.edu |
| Date: 2023-03-28 at 12:00PM |
| Venue: Modern Language 219 |
| Download Flyer Add to Calendar |
| Abstract: Data cleaning and integration is an essential process for ensuring the accuracy and consistency of data used in analytics and decision-making. Schema matching and entity matching tasks are crucial aspects of this process to merge data from various sources into a single, unified view. Schema matching seeks to identify and resolve semantic differences between two or more database schemas whereas entity matching seeks to detect the same real-world entities in different data sources. Given recent deep learning trends, pre-trained transformers have been proposed to automate both the schema matching and entity matching processes. However, existing models only utilize the special token representation (e.g., [CLS]) to predict matches and ignore rich and nuanced contextual information in the description, thereby yielding suboptimal matching performance. To improve performance, we propose the use of the attention mechanism to (1) learn the schema matches between source and target schemas using the attribute name and description, (2) leverage the individual token representations to fully capture the information present in the descriptions of the entities, and (3) jointly utilize the attribute descriptions and entity descriptions to perform both schema and entity matching. |
| Title: Enhancing Document Understanding through the Incorporation of Structural Inference |
|---|
| Defense: Computer Science |
| Speaker: Liyan Xu, Emory University |
| Contact: Dr. Jinho Choi, jinho.choi@emory.edu |
| Date: 2023-03-28 at 1:45PM |
| Venue: Carlos Hall Room 211 |
| Download Flyer Add to Calendar |
| Abstract: Towards resolving a variety of Natural Language Processing (NLP) tasks, pretrained language models (PLMs) have been incredibly successful by simply modeling language sequences, backed by their powerful sequence encoding capabilities. However, for document understanding tasks involving multi-sentence or multi-paragraph inputs, the model still needs to overcome the inherent challenge of processing scattered information across the entire document context, such as resolving pronouns or recognizing relations among multiple sentences. To address the motivation of effectively understanding document context beyond sequence modeling, this dissertation presents an in-depth study on the incorporation of structural inference, utilizing intrinsic structures of languages and documents. Four research works are outlined within this dissertation. Particularly, the first work proposes to integrate syntactic dependency structures into the document encoding process, capturing inter-sentence dependencies through designed graph encoding for the task of machine reading comprehension, especially under the multilingual setting. The second work investigates different methods to perform inference on the discourse structure that concerns coreference relations, allowing for higher-order decision making. The third work presents a novel formulation of structural inference to facilitate joint information extraction, fusing multi-facet information of document entities in terms of both coreference and relations. The last work explores the potential of the sequence-to-sequence generation as an approach that performs implicit inference on linearized entity structures, motivated by its unified encoder-decoder architecture and inherent abilities to perform higher-order inference. Overall, this dissertation demonstrates that incorporating designed structural inference upon certain intrinsic structures of languages or documents can effectively enhance document understanding, and highlights that modeling dependencies among different parts of the context can lead to more accurate and robust encoding and decoding process, where auxiliary information can be provided that complements the sequence modeling of PLMs. Zoom Option: https://emory.zoom.us/j/95753738482 |
| Title: Few Shot Learning for Rare Disease Diagnosis |
|---|
| Seminar: Computer Science |
| Speaker: Emily Alsentzer, MIT & Harvard |
| Contact: Vaidy Sunderam, VSS@emory.edu |
| Date: 2022-12-09 at 1:00PM |
| Venue: https://emory.zoom.us/j/95719302738 |
| Add to Calendar |
| Abstract: Rare diseases affect 300-400 million people worldwide, yet each disease affects no more than 50 per 100,000 individuals. Many patients with rare genetic conditions remain undiagnosed due to clinicians' lack of experience with the individual diseases and the considerable heterogeneity of clinical presentations. Machine-assisted diagnosis offers the opportunity to shorten the diagnostic delays for rare disease patients. Recent advances in deep learning have considerably improved the accuracy of medical diagnosis. However, much of the success thus far is contingent on the availability of large, annotated datasets. Machine-assisted rare disease diagnosis necessitates that approaches learn from limited data and extrapolate beyond training distribution to novel genetic conditions. In this talk, I will present our work towards developing few-shot methods that can overcome the data limitations of deep learning to diagnose patients with rare genetic conditions. To infuse external knowledge into models, we first develop novel graph neural network methods for subgraph representation learning that encode how subgraphs (e.g., a set of patient phenotypes) relate to a larger knowledge graph. We leverage these advances to develop SHEPHERD, a geometric deep learning method that reasons over biomedical knowledge to diagnose patients with rare–even novel–genetic conditions. SHEPHERD operates at multiple facets throughout the rare disease diagnosis process: performing causal gene discovery, retrieving "patients-like-me", and providing interpretable characterizations of novel disease presentations. Our work demonstrates the potential for deep learning methods to accelerate the diagnosis of rare disease patients and has implications for the use of deep learning on medical datasets with very few labels. |
| Title: Language Guided Localization and Navigation |
|---|
| Seminar: Computer Science |
| Speaker: Meera Hahn, |
| Contact: Jinho Choi, jinho.choi@emory.edu |
| Date: 2022-12-02 at 1:00PM |
| Venue: MSC W201 |
| Download Flyer Add to Calendar |
| Abstract: Embodied tasks that require active perception are key to improving language grounding models and creating holistic social agents. In this talk we explore two multi-modal embodied perception tasks which require localization or navigation of an agent in an unknown 3D space with limited information about the environment. First we present the Where Are You? (WAY) dataset which contains over 6k dialogs of two humans performing a localization task. On top of this dataset, we propose the task of Localization from Embodied Dialog (LED). The LED task involves taking a natural language dialog of two agents -- an observer and a locator -- and predicting the location of the observer agent. The second task we examine is the Vision Language Navigation (VLN) task, in which an agent navigates via natural language instructions. For both tasks, we address the objective of improving model accuracy and demonstrate that this can be done using passive data, which can introduce more semantically rich and diverse information during training, in comparison to additional interaction data. We additionally introduce a novel analysis pipeline for both tasks to diagnose and reveal limitations and failure modes of these types of common multi-modal models. Biography: Meera Hahn is a Research Scientist at Google Research working on multi-modal modeling of vision and natural language for applications in artificial intelligence. Her long-term research goal is to develop multi-modal systems capable of supporting robotic or AR assistants that can seamlessly interact with humans. She recently completed her PhD in Computer Science at the Georgia Institute of Technology under Dr. James M. Rehg. Her research at Georgia Tech focused on training embodied agents (in simulation) to perform complex semantic grounding tasks. http://cs.emory.edu/home/ |
| Title: Interpretable and Interactive Representation Learning on Geometric Data |
|---|
| Defense: Computer Science |
| Speaker: Yuyang Guo, Emory University |
| Contact: Liang Zhou, |
| Date: 2022-12-01 at 11:00AM |
| Venue: https://emory.zoom.us/j/5693008550 |
| Download Flyer Add to Calendar |
| Abstract: Abstract: In recent years, representation learning on geometrics data, such as image and graph-structured data, are experiencing rapid developments and achieving significant progress thanks to the rapid development of Deep Neural Networks (DNNs), including Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs). However, DNNs typically offer very limited transparency, imposing significant challenges in observing and understanding when and why the models make successful/unsuccessful predictions. While we are witnessing the fast growth of research in local explanation techniques in recent years, the majority of the focus is rather handling how to generate the explanations, rather than understanding whether the explanations are accurate/reasonable, what if the explanations are inaccurate/unreasonable, and how to adjust the model to generate more accurate/reasonable explanations. To explore and answer the above questions, this dissertation aims to explore a new line of research called Explanation-Guided Learning (EGL) that intervenes the deep learning models' behavior through XAI techniques to jointly improve DNNs in terms of both their explainability and generalizability. Particularly, we propose to explore the EGL on geometric data, including image and graph-structured data, which are currently under-explored in the research community due to the complexity and inherent challenges in geometric data explanation. To achieve the above goals, we start by exploring the interpretability methods for geometric data on understanding the concepts learned by the deep neural networks (DNNs) with bio-inspired approaches and propose methods to explain the predictions of Graph Neural Networks (GNNs) on healthcare applications. Next, we design an interactive and general explanation supervision framework GNES for graph neural networks to enable the learning to explain pipeline, such that more reasonable and steerable explanations could be provided. Finally, we propose two generic frameworks, namely GRADIA and RES, for robust visual explanation-guided learning by developing novel explanation model objectives that can handle the noisy human annotation labels as the supervision signal with a theoretical justification of the benefit to model generalizability. This research spans multiple disciplines and promises to make general contributions in various domains such as deep learning, explainable AI, healthcare, computational neuroscience, and human-computer interaction by putting forth novel frameworks that can be applied to various real-world problems where both interpretability and task performance are crucial. |
| Title: Expressive computation: integrating programming and physical making |
|---|
| Seminar: Computer Science |
| Speaker: Jennifer Jacobs, University of California, Santa Barbara |
| Contact: Vaidy Sunderam, vss@emory.edu |
| Date: 2022-11-18 at 1:00PM |
| Venue: MSC W201 |
| Download Flyer Add to Calendar |
| Abstract: Abstract: Creators in many different fields use their hands. Artists and craftspeople manipulate physical materials, manufacturers manually controlmachine tools, and designers sketch ideas. Computers are increasingly displacing many manual practices in favor of procedural description and automated production. Despite this trend, computational and manual forms of creation are not mutually exclusive. In this talk, I argue that by developing methods to integrate computational and physical making, we can dramatically expand the expressive potential of computers and broaden participation in computational production. To support this argument, I will present research across three categories: 1) Integrating physical and manual creation with computer programming through domain-specific programming environments. 2) Broadening professional computational making through computational fabrication technologies. 3) Broadening entry points into computer science learning by blending programming with art, craft, and design. Collectively, my research demonstrates how developing computational workflows, representations, and interfaces for manual and physical making can enable manual creators to leverage existing knowledge and skills. Furthermore, I’ll discuss how collaborating with practitioners from art, craft, and manufacturing science can diversify approaches to knowledge production in systems engineering and open new research opportunities in computer science. Bio: Jennifer Jacobs is Assistant Professor at the University of California Santa Barbara in Media Arts and Technology and Computer Science (by courtesy). At UCSB, she directs the Expressive Computation Lab, which investigates ways to support expressive computer-aided design, art, craft, and manufacturing by developing new computational tools, abstractions, and systems that integrate emerging forms of computational creation and digital fabrication with traditional materials, manual control, and non-linear design practices. Prior to joining UCSB, Jennifer received her Ph.D. from the Massachusetts Institute of Technology and was a Postdoctoral Fellow at the Brown Institute of Media Innovation within the Department of Computer Science at Stanford University. She also received an M.F.A. and B.F.A from Hunter College and the University of Oregon respectively. Her research has been presented at leading human-computer interaction research venues and journals including UIST, DIS, SIGGRAPH, and, most prominently, at the flagship ACM Conference on Human Factors in Computing Systems (CHI). Zoom Option: https://emory.zoom.us/j/95719302738 |