A generally intelligent agent must be able to teach itself how to solve problems in complex domains with minimal human supervision. Recently, deep reinforcement learning algorithms combined with self-play have achieved superhuman proficiency in Go, Chess, and Shogi without human data or domain knowledge. In these environments, a reward is always received at the end of the game, however, for many combinatorial optimization environments, rewards are sparse and episodes are not guaranteed to terminate. We introduce Autodidactic Iteration: a novel reinforcement learning algorithm that is able to teach itself how to solve the Rubik’s Cube with no human assistance. Our algorithm is able to solve 100% of randomly scrambled cubes while achieving a median solve length of 30 moves — less than or equal to solvers that employ human domain knowledge.

Students: Stephen McAleer, Forest Agostinelli, Alex Shmakov

Advisor: Pierre Baldi, Chancellor’s Professor & Director, Institute for Genomics & Bioinformatics

The ability to observe and analyze the surrounding environment is the key to autonomy. However, advanced data processing is an intense and resource consuming task, especially when contextualized in airborne platforms such as Unmanned Aerial Vehicles (UAV), where weight constraints may limit the availability of vital resources such as energy and processing power. The objective of our research is to enhance the ability of UAVs to process information by effectively distributing computation tasks within UAV swarms and harnessing surrounding IoT resources such as cloud and edge servers.

 Students: Sabur Baidya, Davide Callegaro

Advisor: Marco Levorato, Associate Professor, Computer Science

Cloudberry is a general-purpose middleware system to support visualization on Big Data.  We are developing technologies on approximate query processing, sampling, query slicing, and UI-backend integration to allow users to visualize large amounts of data and gain insights efficiently.  We have a Cloudberry-powered prototype called TwitterMap to support visualization of more than one billion tweets. More information about the project is available at http://cloudberry.ics.uci.edu/.

Student: Qiushi Bai

Advisor: Chen Li, Professor, Computer Science

Automatically building accurate three-dimensional (3D) models of the world from mobile cameras and other sensor data has been a long-term goal in computer vision and robotics research. Many of the key mathematical and computational aspects of this problem have been solved but assume that the world has a rigid, static geometry. This severely limits the applicability of such methods in dynamic environments, such as a construction site, where the “map” of the world can change radically from one day to the next. We are developing techniques that take into account changes over time, allowing one to construct 4D maps of the world from streams of camera data.

Student: Minhaeng Lee

Advisor: Charless Fowlkes, Professor, Computer Science

The TIPPERS project, funded by DARPA through the Brandeis program, explores the implications of the Internet of Things (IoT) in our privacy. We are developing a system that captures IoT data, transforms it into higher-level abstractions, and offers it to developers to build IoT applications. The system implements a privacy-by-design approach in which different privacy technologies such as policies, differential privacy, or secure computing, are implemented in the different phases of IoT data management. We have created a testbed based on the TIPPERS system in our Donald Bren Hall building, and are now expanding the service to other buildings in the UCI campus, to allow researchers on privacy preserving technologies to test their technologies in live IoT scenarios.

Students: Abdulrahman Alsaudi, Arda Unal, Daokun Jiang, Dhrubajyoti Ghosh, Eun-Jeong Shin, Guoxi Wang, Peeyush Gupta, Primal Pappachan, Rushabh Shah, Sameera Ghayyur, Sumaya Abdullah A Almanee, Yiming Lin

Advisors: Sharad Mehrotra, Professor, Computer Science; Nalini Venkatasubramanian, Professor, Computer Science; Alfred Kobsa, Professor, Informatics; Roberto Yus, Postdoctoral Scholar, Computer Science; Shantanu Sharma, Postdoctoral Scholar, Computer Science; Nisha Panwar, Postdoctoral Scholar, Computer Science

Take control of a VR conversation between magical sisters about the fate of their fairy tale world. In this collaboration with Broadway producer Tim Kashani, we create a novel new participatory theater experience in the world of one of his upcoming productions, The Next Fairy Tale. The interactor is cast as Caliope, and must decide how the emotional experience of a scene plays out by casting spells during a conversation with Minerva: the villain of the show. Spells cast by the player impact the vocal performance of the characters, and the emotional context of the script, resulting in a unique and personal version of the scene for each player. This piece explores new forms of interacting with digital stories in virtual reality by allowing virtual performers to make expressive choices similar to those made by actors on a stage.

Students: Mengfan Wang, Nicholas Persa, Minnie Wu, Saumya Gupta, Meena Muralikumar, Parker Scott, Ace Lowder, Edward Lok

Advisor:  Joshua Tanenbaum, Assistant Professor, Informatics

Analytics with Big Data has become a very challenging task today due to the growing complexity of handling massive datasets with various toolkits. Data Analysts are forced to deal with different environments instead of focusing on analyzing the data. In this project, we introduce a data analytic library (AFrame) that we build on top of Apache AsterixDB. This library allows users to interact with a large amount of distributed data in the same way that Pandas data frame (Python analytic library) works with locally stored data. We integrate Pandas with AsterixDB to embrace analysts with their familiar environment, and scale-out the evaluation of the analytical operations over a distributed database system to enable real Big Data Analytics. With this new framework, we hope we can make Big Data Analytics more accessible and efficient for our users.

Students: Xikui Wang, Phanwadee Sinthong

Advisors: Michael Carey, Bren Professor, ICS; Chen Li, Professor, Computer Science

Achieving positive health and fitness outcomes is a challenge for the visually impaired community. Limited opportunities, dependence on sighted guides, and expensive assistive devices pose significant barriers to most outdoor activities. In collaboration with the non-profit organization Makapo, we have developed a low-cost remote control harness for one person rudder controlled boats. Our harness enables blind and low-vision paddlers to experience independence on the water, engage in healthy outdoor activity, and participate in a social environment.

Student: Mark S. Baldwin

Advisors: Gillian R. Hayes, Robert A. and Barbara L. Kleist Professor, Informatics; Jennifer Mankoff

Colorectal cancer (CRC)  is the second leading cause of cancer related death in the U.S. CRC arises from precancerous polyps with a mean dwell time of 10+ years. The National Polyp Study showed that 70%-90% of colorectal cancers are preventable with regular colonoscopies and removal of polyps. Seven to nine percent of colorectal cancers occur despite being up-to-date with colonoscopy. It is estimated that 85% of these “interval cancers” are due to missed polyps or incompletely removed polyps during colonoscopy. This projects develops and applies AI and machine learning (deep learning) methods to the problem of automatically detecting polyps in colonoscopy videos to improve detection rates. The trained system achieves an accuracy of 96% and can be deployed in real time. Developed in collaboration with Dr. W. Karnes and the UCI Department of Gastroenterology.

Student: Gregor Urban

Advisor: Pierre Baldi, Distinguished Professor & Director, Institute for Genomics & Bioinformatics

Knowledge bases (KB) are an essential part of many computational systems with applications in search, structured data management, recommendations, question answering, and information retrieval. However, KBs often suffer from incompleteness, noise in their entries, and inefficient inference under uncertainty. To address these issues learning relational KBs by representing entities and relations in an embedding space has been a focus of active research. However, knowledge bases in the real-world, contain a much wider variety of data types such as text, images, and numerical values, which are being ignored by current methodology. We propose multimodal knowledge base embeddings (MKBE) that use different neural encoders for this variety of observed data, and combine them with existing relational models to learn embeddings of the entities and multimodal data. Further, using these learned embeddings and different neural decoders, we introduce a novel imputation model to generate missing multimodal values, like text and images, from information in the knowledge base.

Student: Pouya Pezeshkpour, Liyan Chen

Advisor: Sameer Singh, Assistant Professor, Computer Science

Deep learning has experienced an explosion of popularity due to its power and flexibility across a wide range of problems and data.  However, due to the small sample sizes for typical neuroimaging studies, deep learning has only seen limited applications in this area.  We here present a deep learning model that utilizes the graphical structure of brain connectivity to predict the experimental condition of a subject during a gambling task (win/loss domain) from their fMRI data.  This model exhibits substantially better predictive accuracy relative to traditional neural networks and other machine learning methods, suggesting that the relationship of brain regions is important for discriminating the experimental domains.

Students: Dustin Pluta, Lingge Li

Advisors: Zhaoxia Yu, Vice Chair, Undergraduate Affairs & Associate Professor, Statistics;  Hernando Ombao, Professor, Statistics; Pierre Baldi, Chancellor’s Professor & Director, Institute for Genomics & Bioinformatics; Chuansheng Chen, Chair & Professor, Psychological Science; Babak Shahbaba, Vice Chair Graduate Affairs & Associate Professor, Statistics

In epidemiological studies, disease risk patterns over geographical space are of common interest since spatial variations in risk are potential indications of latent risk factors and health disparities, such as water pollution or availability of sufficient health care. We develop new statistical models to flexibly estimate time-specific disease risk patterns. Further, we propose a data re-sampling strategy to formally test the temporal heterogeneity of spatial risk patterns. We apply our methods to data on the incidence of patent ductus arteriosus (a rare, but serious, birth defect) in the state of Massachusetts and illustrate heterogeneity in spatial risk patterns over the years 2003, 2006 and 2009. This work is performed in collaboration with Professors Veronica Vieira and Scott Bartell, UCI Program in Public Health.

Student: Yannan Tang

Advisor: Dan Gillen, Chair & Professor, Statistics

Current VR/AR technology requires clunky wearable devices that inhibit natural human interaction. Here we show an interactive multi-projector display that turns any physical object of any scale or shape into a display with which one or more users can interact using different interaction modalities like hand gestures, laser pointers or mobile devices. This enables creation of unique shareable 3D experiences. 

Students: Mahdi Abbaspour Tehrani, Mehdi Rahimzadeh, Twaha Ibrahim

Advisors: Aditi Majumder, Professor, Computer Science; Gopi Meenakshisundaram, Associate Dean, ICS Student Affairs & Professor, Computer Science

The research and development of wearable technologies for self-regulation of attention.

Students: Craig Anderson, Nicholas Persa, Richard Martinez, Maria Anderson Soto, Max Collins, Richie Poon, Zachary Cloutier, Caleb Chu, Jackson Greaves

Advisor: Kurt Squire