Colloquia 2018-2019

Tuesday, February 19, 2019.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Jeff Clune, PhD

Title: Understanding and Improving Deep Neural Networks

Abstract:

With deep learning, deep neural networks have produced state-of-the-art results in a number of different areas of machine learning, including computer vision, natural language processing, robotics and reinforcement learning. I will summarize three projects on better understanding deep neural networks and improving their performance. First I will describe our sustained effort to study how much deep neural networks know about the images they classify. Our team initially showed that deep neural networks are “easily fooled,” meaning they will declare with near certainty that completely unrecognizable images are everyday objects, such as guitars and starfish. These results suggested that deep neural networks do not truly understand the objects they classify. However, our subsequent results reveal that, when augmented with powerful priors, deep neural networks actually have a surprisingly deep understanding of objects, which enables them to be incredibly effective generative models that can produce a wide diversity of photo-realistic images. Second, I will summarize our Nature paper on learning algorithms that enable robots, after being damaged, to adapt in 1-2 minutes and soldier on with their mission. This work combines a novel stochastic optimization algorithm with Bayesian optimization to produce state-of-the-art robot damage recovery. Third, I will describe our recent Go-Explore algorithm, which dramatically improves the ability of deep reinforcement learning algorithms to solve previously unsolvable problems wherein reward signals are sparse, meaning that intelligent exploration is required. Go-Explore solves Montezuma’s Revenge, considered by many to be a grand challenge of AI research. I will also very briefly summarize a few other machine learning projects from my career, including our PNAS paper on automatically identifying, counting, and describing wild animals in images taken remotely by motion-sensor cameras.

Bio:

Jeff Clune is the Loy and Edith Harris Associate Professor in Computer Science at the University of Wyoming and a Senior Research Manager and founding member of Uber AI Labs, which was formed after Uber acquired a startup he helped lead. Jeff focuses on robotics and training neural networks via deep learning, including deep reinforcement learning. Since 2015, a robotics paper he co-authored was on the cover of Nature, a deep learning paper from his lab was on the cover of the Proceedings of the National Academy of Sciences, he won an NSF CAREER award, he received the Distinguished Young Investigator Award from the International Society for Artificial Life, he was an invited speaker at the NeurIPS Deep Reinforcement Learning Workshop, and his deep learning papers were awarded honors (best paper awards and/or oral presentations) at the top machine learning conferences (NeurIPS, CVPR, ICLR, and an ICML workshop). His research is regularly covered in the press, including the New York Times, NPR, NBC, Wired, the BBC, the Economist, National Geographic, the Atlantic, the New Scientist, the Daily Telegraph, Science, Nature, and U.S. News & World Report. Prior to becoming a professor, he was a Research Scientist at Cornell University, received degrees from Michigan State University (a PhD and a master’s degree) and the University of Michigan (a bachelor’s degree). More information about Jeff's research is available at http://JeffClune.com.

Host: Mihai Surdeanu

 

Tuesday, February 12, 2019.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Jason Pacheco, PhD

Title: Probabilistic Reasoning in Complex Systems: Algorithms and Applications

Abstract:

Statistical machine learning approaches to scientific applications are complicated by high-dimensional, continuous, and nonlinear interactions that typically arise in such settings. In this talk I will discuss several general-purpose approaches to reasoning in these systems. I will introduce Diverse Particle Max-Product (D-PMP), a particle-based extension of max-product belief propagation (BP) for maximum a posteriori (MAP) inference and will show how D-PMP isolates multiple distinct, locally-optimal, configurations through a diverse particle selection process. I will further demonstrate how D-PMP can be flexibly adapted to problems of articulated human pose estimation in images and video, as well as protein structure prediction from low-resolution experimental data. In the second half of the talk I will present robust and efficient algorithms for sequential decision making in information gathering systems. These algorithms sequentially choose among actions to maximally reduce uncertainty about quantities of interest. Using gene regulatory network inference as an example, I will demonstrate how these approaches meet or exceed the performance of domain-specific methods for Bayesian experimental design.

Bio:

Jason Pacheco is a postdoctoral associate at MIT in the Computer Science and Artificial Intelligence Laboratory (CSAIL) with John Fisher III. Prior to joining MIT Jason completed his graduate work at Brown University with Erik Sudderth. Jason’s research interests are in statistical machine learning, probabilistic graphical models, approximate inference algorithms, and information-theoretic decision making.

Host: Kobus Barnard

 

Thursday, February 07, 2019.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: David Inouye, PhD

Title: Deeper Understanding of Deep Learning via Shallow Learning and Model Explanations

Abstract:

Despite tremendous empirical success, modern deep learning is still relatively new, and thus, there are significant gaps in understanding compared to classical shallow learning. This lack of deep understanding of deep learning hinders practitioners from systematically and reliably developing models especially in new contexts—thus development is often relegated to laborious trial and error. In addition, this lack of understanding hinders users from adopting deep models in real-world applications. As one approach for deeper understanding, I will discuss how to leverage well-understood shallow learning to construct deep models so that the algorithms and insights from shallow learning can be lifted into the deep context. Specifically, I will present a destructive process that iteratively finds patterns in the data via shallow learning and then destroys these patterns via transformations. I will then show an application of this unconventional deep learning approach to deep probabilistic models. As a different approach, model explanations can increase users' understanding of deep models and thereby aid them in deciding if they should adopt a deep model or not. Thus, I will also describe how to create model explanations based on the concept of counterfactuals that are simultaneously exact and non-local—in contrast to explanations based on local approximations of the model. I will present a new framework for finding useful model explanations, conceptualized as lines or curves in the input space that compress the full model into a few relevant trajectories for a given target point, where relevancy depends on the context. In conclusion, I will discuss promising future directions for both destructive learning and model explanation.

Bio:

David Inouye is a postdoctoral researcher at Carnegie Mellon University in the Machine Learning Department working with Prof. Pradeep Ravikumar. He completed his PhD in CS at The University of Texas at Austin where he was advised by Prof. Inderjit Dhillon and Prof. Pradeep Ravikumar. David's research interests include deep generative models, probabilistic graphical models, model explanations and model visualizations. He completed his BS in EE at Georgia Institute of Technology and was awarded the NSF GRFP graduate research fellowship during his senior year.

Host: Carlos Scheidegger

 

Tuesday, January 22, 2019.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Cynthia Bailey, PhD

Title: What can I do today to create a more inclusive community in CS?

Abstract:

Many people who work in STEM wish that our field were more diverse and earnestly want to be part of the solution, yet aren't sure where to begin. This talk will focus on concrete, actionable suggestions that anyone in our field can implement to help create a more inclusive and equitable environment. Drawing on lessons learned in her teaching, research, and mentorship programs, Dr. Lee will provide an inclusion toolkit ranging from 5-minute targeted interventions to lifelong habits.

Bio:

Cynthia Lee is a Lecturer in the Computer Science Department at Stanford University. Her research focuses on education and social impacts of technology. She founded peerinstruction4cs.org to assist faculty in redesigning their CS coursework around research-based best practices. Her work experience includes NASA Ames and startups in the machine learning, children’s education apps, and augmented reality spaces. She holds a PhD from University of California, San Diego where her research focused on machine learning and market-based algorithms for high performance computing systems. She was voted the 2015 Professor of the Year by Stanford Society of Women Engineers.

Host: Michelle Strout

 

Tuesday, January 15, 2019.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Michael Gleicher, PhD

Title: Interpreting Embeddings with Comparison

Abstract:

Vector embeddings place objects (e.g., documents or words) in a vector space such that similar objects are close. Embeddings can abstract information from data collections and have wide usage in fields such as machine learning, and natural language processing. However, embeddings are challenging to interpret, which limits the use of embeddings as a tool for understanding the underlying data or the performance of the methods that construct embeddings. In this talk, I will survey our work in building visualization tools that address challenges in interpreting embeddings. I will use the idea of comparison as a strategy for designing data analysis tools. I will introduce our framework for thinking about comparison, and show how we used this framework in the design of tools for embedding challenges. I will provide examples of tools for examining and comparing both document embeddings (e.g., topic models) and word vector embeddings.

Bio:

Michael Gleicher is a Professor in the Department of Computer Sciences at the University of Wisconsin, Madison. Prof. Gleicher is founder of the Department's Visual Computing Group. His research interests span the range of visual computing, including data visualization, robotics, image and video processing tools, virtual reality, and character animation. His current foci are human data interaction and human robot interaction. Prior to joining the university, Prof. Gleicher was a researcher at The Autodesk Vision Technology Center and in Apple Computer's Advanced Technology Group. He earned his Ph. D. in Computer Science from Carnegie Mellon University, and holds a B.S.E. in Electrical Engineering from Duke University. In 2013-2014, he was a visiting researcher at INRIA Rhone-Alpes. Prof. Gleicher is an ACM Distinguished Scientist.

Host: Carlos Schiedegger

 

Wednesday, December 05, 2018.

Time: 11:00am

Location:  Gould Simpson (GS) 701

Speaker: Patrick Verga

Title: Extracting and Embedding Entities, Types, and Relations

Abstract:

Over the past five years, deep learning has shown remarkable success on shallow natural language processing (NLP) tasks and in perceptual domains like vision. This has large been driven by models which largely ignore explicit representations of knowledge in favor of unconstrained, end-to-end learning over massive supervised training sets. On shallow perceptual tasks this is typically adequate, but as we seek to develop methods for true language understanding, knowledge representation and reasoning will be essential and the question of how best to represent and acquire knowledge remains open.

In this talk I will present methods for incorporating powerful neural network models with rich structured information to improve the representations of entities and their relations while extracting new knowledge from raw unstructured text. Symbolic knowledge graphs over fixed, human-defined schema encode facts about entities and their relations but are brittle, lack specificity, and are highly incomplete. Universal schema (US) addresses all of these issues by learning a latent schema over both existing structured resources and unstructured free text data, embedding them jointly within a shared space. We first improve generalization in US by (1) representing text using a compositional neural encoder, leading to substantially improved recall and enabling zero-shot relation extraction in a language with no training data, and (2) representing entities and entity pairs as query-specific aggregations over observed textual evidence rather than a static global embedding. We also improve accuracy in fine-grained entity typing and linking by injecting additional structure into our embedding space, enforcing a hypernym hierarchy over entities and types. Finally, we present a model for extracting all entities and relations simultaneously over full paragraphs, improving biological relation extraction by allowing for finer-grained encodings and extraction of long range cross-sentence relations.

Bio:

Patrick Verga is a final year PhD candidate in the College of Information and Computer Sciences at UMass Amherst, advised by Andrew McCallum. His research contributes to knowledge representation and reasoning, with a focus on large knowledge base construction from unstructured text, with applications to general domain, commonsense, and biomedicine. Pat previously interned at Google and the Chan Zuckerberg Initiative and received a best paper award at EMNLP 2018. Over the past several years he has advised multiple M.S. and junior PhD students, resulting in published research in fine-grained entity typing linking, unsupervised parsing, and partially labeled named entity extraction. He holds M.S. and B.A degrees in computer science as well as a B.S. in neuroscience.

Host: Kobus Barnard

 

Tuesday, December 04, 2018.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Emma Strubell

Title: Neural Network Architectures for Efficient and Robust NLP

Abstract:

NLP has come of age. For example, semantic role labeling (SRL), which automatically annotates sentences with a labeled graph representing "who" did "what" to "whom," has in the past ten years seen nearly 40% reduction in error, bringing it to useful accuracy. As a result, hoards of practitioners now want to deploy NLP systems on billions of documents across many domains. However, state-of-the-art NLP systems are typically not optimized for cross-domain robustness nor computational efficiency.

In this talk I will present two new methods to facilitate fast, accurate and robust NLP. First, I will describe Iterated Dilated Convolutional Neural Networks (ID-CNNs, EMNLP 2017), a faster alternative to bidirectional LSTMs for sequence labeling, which in comparison to traditional CNNs have better capacity for large context and structured prediction. Unlike LSTMs whose sequential processing on sentences of length N requires O(N) time even in the face of GPU parallelism, ID-CNNs permit fixed-depth convolutions to run in parallel across entire documents. They embody a distinct combination of network structure, parameter sharing and training procedures that enable dramatic 14-20x test-time speedups while retaining accuracy comparable to the Bi-LSTM-CRF. Second, I will present Linguistically-Informed Self-Attention (LISA, EMNLP 2018 Best Long Paper), a neural network model that combines multi-head self-attention with multi-task learning across dependency parsing, part-of-speech tagging, predicate detection and SRL. Unlike previous models which require significant pre-processing to prepare syntactic features, LISA can incorporate syntax using merely raw tokens as input, encoding the sequence only once to simultaneously perform parsing, predicate detection and role labeling for all predicates. Syntax is incorporated through the attention mechanism, by training one of the attention heads to focus on syntactic parents for each token. We show that incorporating linguistic structure in this way leads to substantial improvements over the previous state-of-the-art (syntax-free) neural network models for SRL, especially when evaluating out-of-domain, where LISA obtains nearly 10% reduction in error while also providing speed advantages. 

Bio:

Emma Strubell is a final-year PhD candidate in the College of Information and Computer Sciences at UMass Amherst, advised by Andrew McCallum. Her research aims to provide fast, accurate, and robust natural language processing to the diversity of academic and industrial investigators eager to pull insight and decision support from massive text data in many domains. Toward this end she works at the intersection of natural language understanding, machine learning, and deep learning methods cognizant of modern tensor processing hardware. She has applied her methods to scientific knowledge bases in collaboration with the Chan Zuckerberg Initiative, and to advanced materials synthesis in collaboration with faculty at MIT. Emma has interned as a research scientist at Amazon and Google and received the IBM PhD Fellowship Award. She is also an active advocate for women in computer science, serving as leader of the UMass CS Women’s group where she co-organized and won grants to support cross-cultural peer mentoring, conference travel grants for women, and technical workshops. Her research has been recognized with best paper awards at ACL 2015 and EMNLP 2018.

Host: Kobus Barnard

 

Thursday, November 29, 2018.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Paul Medvedev

Title: Assembly of big genomic data

Abstract: 

As genome sequencing technologies continue to facilitate the generation of large datasets, developing scalable algorithms has come to the forefront as a crucial step in analyzing these datasets. In this talk, I will discuss several recent advances, with a focus on the problem of reconstructing a genome from a set of reads (genome assembly). I will describe low-memory and scalable algorithms for automatic parameter selection and de Bruijn graph compaction, recently implemented in two tools KmerGenie and bcalm. I will also present recent advances in the theoretical foundations of genome assemblers.

Bio: 

Paul Medvedev is an Associate Professor in the Department of Computer Science and Engineering and the Department of Biochemistry and Molecular Biology and the Director of the Center for Computational Biology and Bioinformatics at the Pennsylvania State University. His research focus is on developing computer science techniques for analysis of biological data and on answering fundamental biological questions using such methods. Prior to joining Penn State in 2012, he was a postdoc at the University of California, San Diego and a visiting scholar at the Oregon Health & Sciences University and the University of Bielefeld. He received his Ph.D. from the University of Toronto in 2010, his M.Sc. from the University of Southern Denmark in 2004, and his B.S. from the University of California, Los Angeles in 2002. 

Host: John Kececioglu

 

Tuesday, November 27, 2018.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Weihao Kong

Title: The surprising power of little data

Abstract:

Despite the rapid growth of the size of our datasets, the inherent complexity of the problems we are solving is also growing, if not at an even faster rate. This prompts the question of how to infer the most information from the available data. 

I will discuss several examples of my research that reveal a surprising ability to extract accurate information from modest amounts of data. The first setting that I discuss considers data provided by a large number of heterogeneous individuals, and we show that the empirical distribution of the data can be significantly "de-noised". The second setting considers estimating the intrinsic dimensionality of a dataset, in the sublinear sample regime where the empirical distribution of the data is misleading. The final portion of my talk focuses on estimating "learnability": given too little data to learn an accurate prediction model, we can accurately estimate the value of collecting more data. Specifically, for some natural model classes, we can estimate the performance of the best model in the class, given too little data to find any model in the class that would achieve good prediction error. In most of these settings, our algorithms are provably information-theoretically optimal and are also highly practical.

Bio:

Weihao Kong is a Ph.D. student from the Computer Science Department at Stanford University. He received his B.S. from Shanghai Jiao Tong University in 2013. His research interests span machine learning, high-dimensional statistics and theoretical computer science.

Host: Carlos Scheidegger

 

Tuesday, November 15, 2018.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: ALFRED Z. SPECTOR, CTO at Two Sigma 

Title: Research Challenges in Computer Science

Abstract:

The trillion-fold increase in the capability of computation over the past 60 years, when coupled with global connectivity and vast amounts of available data makes, for a vibrant field that will continue to grow and provide many research and employment opportunities for computer scientists. In this talk, I describe a plethora of research challenges, including but not limited to the grand challenge problems of Artificial Intelligence. I’ll discuss research objectives relating to
scalability, usability, security, robustness, knowledge representation and inferencing, application areas such as healthcare and education, and more. I’ll discuss the breadth of challenges and the diversities of talent that will be required to meet them. I’ll conclude with
some example approaches to these problems from my experience leading research teams in academia and industry.

Bio:

Alfred Spector is Chief Technology Officer at Two Sigma, a firm dedicated to using information to undertake many forms of economic optimization. Dr. Spector's career has led him from innovation in large scale, networked computing systems (at Stanford, CMU, and his company, Transarc) to broad research leadership: eight years leading Google Research and five years leading IBM Software Research. Recently, Spector has lectured widely on the growing importance of computer science across all disciplines (CS+X) and on the Societal Implications of Data Science. He received an AB in Applied Mathematics from Harvard and a Ph.D. in Computer Science from Stanford. He is a Fellow of the ACM and IEEE, and a member of the National Academy of Engineering and the American Academy of Arts and Sciences. Dr. Spector won the 2001 IEEE Kanai Award for Distributed Computing and was co-awarded the 2016 ACM Software Systems Award.

Host: Tom Fleming, Astronomer and Senior Lecturer in Astronomy

 

Tuesday, November 06, 2018.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Hong Hu, PhD

Abstract:

Memory corruption bugs, like buffer overflow or use-after-
free, enable attackers to manipulate victim's memory for 
their bidding. Due to the lack of complete and efficient 
bug detection system, attackers will still have exploitable 
memory bugs in the near future. Unfortunately, in current 
competition between attackers and defenders, the latter 
always take actions passively after new attacks and 
bypasses happen, leading to ineffective protection.

In this talk, I will present two of my work that help regain 
the initiative in the competition on memory bugs. In the 
first work, we develop a control-flow protection 
mechanism that prevents any control-flow hijacking 
attacks, with the purpose of ending the long-competition 
between control-flow integrity proposals and bypasses. 
In the second work, we propose a new attack 
methodology to explore the expressiveness of data-only 
attacks. We show that without corrupting any control-flow 
attackers are able to cause severe damage, evening 
achieving Turing-complete computing in the victim's 
memory. These work shows that partial memory safety 
cannot guarantee the program security, and we should 
spend more effort on developing efficient full memory 
safety mechanisms.

Bio:

Dr. Hong Hu is a postdoctoral fellow in the School of 
Computer Science, College of Computing, the Georgia 
Institute of Technology. His main research area is system 
security, focusing on detecting memory errors from 
C/C++ programs, exploring new attack vectors and 
developing defense mechanisms to prevent exploits. His 
work has appeared in top venues in system security 
(USENIX Security, IEEE Security and Privacy, CCS, ICECCS 
and ESORICS). He received the Best Paper Award from 
ICECCS 2014. He obtained his Ph.D. degree in computer 
science from the National University of Singapore in 
2016, and his B.E. degree in information security from 
the Huazhong University of Science and Technology in 2011.

Host: Saumya Debray

 

Thursday, October 04, 2018.

Time: 11:00am

Location:  Gould Simpson (GS) 906

Speaker: Daniel Fried, PhD

Title: Pragmatic Models for Generating and Following Grounded Instructions

Abstract:

A system that interacts with people in natural language predicts what
it should say -- why not also predict how the person listening will
react? We describe methods for generating and following natural
language instructions by incorporating pragmatics: explicitly modeling
people and world contexts. Our pragmatics-enabled models reason about
how listeners will carry out instructions, and reason counterfactually
about why speakers produced the instructions they did. We find that
this reasoning procedure improves state-of-the-art listener models (at
correctly following human instructions) and speaker models (at
generating instructions correctly interpretable by humans) for
sequential tasks in diverse settings, including navigating through
real-world indoor environments.

Bio:

Daniel Fried is a PhD student at UC Berkeley working on natural
language processing and machine learning, with a focus on grounded
semantics and structured prediction. Previously, he received a BS from
the University of Arizona and an MPhil from the University of
Cambridge. His work has been supported by a Churchill Scholarship,
NDSEG Fellowship, Huawei / Berkeley AI Fellowship, and Tencent
Fellowship.

Host: Stephen Kobourov

 

Tuesday, August 14, 2018.

Time: 11:30am

Location:  Gould Simpson (GS) 701

Speaker: Ahyoung Choi, Assistant Professor, Gachon University, Republic of Korea

Title: Health sensing by wearables and its application to obesity management and vital sign monitoring

Abstract: 

Precision medicine is an emerging approach to focus on prevention and treatment considering individual differences in genes and lifestyles rather than focusing on diagnosing the disease in a uniform way by population-based statistics and science. This approach accelerates the health monitoring and wellness monitoring device market and its integration of traditional medical records with the viewpoint of big data. For example, GearFit, Fitbit, Jawbone, G Watch Urbane and other wearable devices monitor daily exercise activity, sleep quality, and dietary habits for wellness monitoring. Portable blood pressure monitors and patch-type ECG measurement sensors collect health-related data in an unobtrusive way in daily life for health monitoring. However, such health-sensing wearable devices and methods have not yet been used as a substitute for digital healthcare widely because of low usability as well as low accuracy. This talk introduces health sensing and analysis method using wearable devices in everyday life while ensuring usability in obesity management and vital sign monitoring.

Bio:

Ahyoung Choi is currently an assistant professor in the department of software at Gachon University (Seongnam, Republic of Korea). She received her M.S. and Ph.D. in the department of information and communications at Gwangju Institute of Science and Technologies (Gwangju, Republic of Korea) in 2005 and 2011, respectively. She was a visiting scholar at Institute for Creative Technologies at University of Southern California (CA, USA) from 2011 to 2012 and worked at Samsung Electronics (Suwon, Republic of Korea) from 2012 to 2016. Her research interests include physiological signal processing, and human-computer interaction and its application to mobile healthcare systems.

Host: Stephen Kobourov