Vashistha, now a computing and information science professor at Cornell University, completed his Ph.D. working with professor Richard Anderson in the Information & Communication Technology for Development (ICTD) Lab on social computing technologies for underserved communities in low-resource environments. His dissertation focused on designing, building and evaluating new computer technologies to include people who are often excluded from social computing platforms because they are too poor to afford smartphones, too remote to access the internet or too low-literate to navigate all the text on the internet.
“Being born and raised in India where illiteracy, poverty, and social ills were just a stone’s throw away, I witnessed how non-reading people, low-income women, and people with disabilities struggled with digital inequity and social injustice,” Vashistha said. “These experiences have profoundly motivated me to build computing technologies that include people of all backgrounds in the information revolution, particularly marginalized communities who are often neglected by the designers and builders of mainstream technologies.”
The first step in his journey began while working with Bill Thies of Microsoft India. Vashistha created IVR Junction, a system that uses interactive voice response (IVR) technology to enable people with basic phones to participate in voice-based social networks. Deployed in remote regions of Somalia, Mali and India, IVR Junction gave communities a way to share news, call attention to rights violations and report lack of services. Given the explosive growth of these systems, Vashistha then focused on Sangeet Swara, an interactive voice forum that enables people in these rural areas to moderate and manage the content generated in local languages which are yet unsupported by advances in natural language processing. The paper presenting Sangeet Swara earned a Best Paper Award at the 2015 ACM Conference on Human Factors in Computing Systems. Vashistha subsequently earned an Honorable Mention at the CHI 2017 for his work on Respeak, a voice-based speech transcription that relies on crowd-sourcing and speech recognition to transcribe audio files while providing additional earning opportunities to low-literate people without access to smartphones and internet connectivity.
In addition to building voice forums for all people, Vashistha studied how technology amplifies existing sociocultural norms and values in society, including its strengths, shortcomings and biases. For example, he found that while Sangeet Swara transformed the lives of low-income blind people in rural regions, it also exposed that women who were marginalized due to patriarchy-driven abuse and hate speech. His dissertation advances the discourse on the benefits and pitfalls of social computing, highlights new challenges and big frontiers in building social good applications in low-resource environments, and offers solutions to make computing technologies more diverse, inclusive, and impactful.
Both the Levinson Scholars, funded by Art (UW ‘72) and Rita Levinson, and the WRF Fellows are chosen for their innovative research in bioscience and other related fields. The scholarships will enable them to pursue their projects while continuing to be supported by their mentors and lab colleagues.
Skyler Hallinan
Hallinan is a senior majoring in computer science, bioengineering, and applied and computational mathematical sciences. He works with bioengineering professor Paul Yager.
Hallinan’s research focuses on combating chronic kidney disease (CKD), a disease that affects millions globally. People with CKD often accumulate indoxyl sulfate, a uremic toxin normally filtered out by healthy kidneys, which can cause major illness and lead to death. Hallinan is working to develop a method to effectively remove indoxyl sulfate from CKD patients’ blood via an orally ingestible hydrogel. He is currently modeling, prototyping and testing different candidate hydrogels to engineer a substitute for a functional kidney.
Karim’s research focuses on measuring stress levels in teens and creating therapeutic, intervening techniques for them with a social robot called EMAR (Ecological Momentary Assessment Robot). Using EMAR, Karim collects stress level measurement data from high school students, and offers interventions that draw from dialectic behavioral therapy and acceptance and commitment therapy. Through this robot interaction experience, Karim aims to help teens be more mindful and present in the moment while developing a robot with a heightened level of sensitivity.
Ximing Lu
Lu is a computer science and statistics major working with professor Linda Shapiro, who holds a joint appointment in the Allen School and Department of Electrical & Computer Engineering, and with Allen School professor Yejin Choi as a research intern at the Allen Institute for AI.
Lu’s current research with Choi centers around natural language processing and commonsense reasoning. With Shapiro, she focuses on creating a computer-aided biopsy classification system to reduce cancer diagnosis uncertainties. The system, HATNet: An End-to-End Holistic Attention Network for Diagnosis of Cancer Biopsy Images, streamlines the histopathological image classification pipeline and shows how to learn representations from gigapixel size images end-to-end. HATNet can learn representations from clinically relevant tissue structures and match the classification accuracy of pathologists.
Cao is currently working on the development of a non-invasive, continuous blood pressure monitor to help patients with cardiovascular conditions. In particular, his work aims to help people with postural orthostatic tachycardia syndrome (POTS), a condition where the body cannot properly regulate blood vessels that causes lightheadedness, fainting, and spikes in heart rate. The device will collect the pulse transit time (PTT), the time for a pulse wave to travel between two points. PTT correlates with blood pressure, which is a known predictor of adverse POTS symptoms. In addition to helping predict symptoms, the device will also provide physicians with a dataset that will make conditions like POTS easier to diagnose.
Jakup Filipek
Filipek is a senior computer science major enrolled in the Allen School’s B.S./M.S. program. He works with physics professor Shih-Chieh Hsu.
Filipek’s research is in quantum machine learning (QML). Quantum computers have the potential to produce results in simple artificial intelligence algorithms to sophisticated neural networks better than their classical counterparts. However, these models are dealing with the bottleneck issue of a limited number of quidbits (the basic unit of quantum information) in near-term quantum devices. FIlipek is working on a hybrid neural network that functions by sandwiching any QML algorithm between two classical neural networks. This allows for automatic scaling of quantum algorithms to inputs and outputs of any size, addressing the bottleneck issue while provisioning an easy way of comparing classical algorithms to quantum ones.
Millicent Li
Li is a senior in computer science working with Patel in the Ubiquitous Computing Lab.
Li’s research is in improving the communication abilities of people with speech impairments by developing a silent speech interface that can facilitate communication between two people or with smart devices by outputting speech that is imagined but not spoken. Li aims to use a combination of neural signals from the brain to sense the user’s intent in their day-to-day lives and to provide speech accordingly, without the need of facial movement, by using optical signals like functional near-infrared spectroscopy and electrical signals like electroencephalogram.
Congratulations to these outstanding Allen School scholars who are making an impact through undergraduate research!
Our latest undergraduate student spotlight features Sammamish, Washington native Joe Spaniac, a second year computer science major admitted under the Allen School’s expanded Direct to Major admission pathway who is also majoring in drama. During his first quarter at the Allen School, he knew he wanted to be a part of Q++, a student organization for LGBTQIA+ members studying in the Allen School. He has served as a member of the board since then and currently co-chairs the group with Lavinia Dunagan. Despite the challenges of meeting and holding events virtually this year, Spaniac enjoys being a part of the group and all that it does for students in the Allen School.
Allen School: What is the mission of Q++?
Joe Spaniac: Overall, our mission with Q++ is to raise awareness of the issues that LGBTQIA+ members face in the field of computer science and provide a place that anyone and everyone can feel welcome and supported regardless of gender expression or sexual orientation. This year especially, we are really focused on the second half of that mission statement due to the many social challenges that come with, seemingly, an entirely online school year. Very frequently, members of the community really look forward to the typical “college experience” as an opportunity to reinvent themselves and sometimes even escape households that might not accept them as their true selves. Although Q++ may not be able to entirely replace this experience, we hope to be a good intermediate option while we all wait for things to return to normal.
Allen School: Why do you think it’s important to have an organization like Q++ at the Allen School?
JS: Speaking from personal experience, the UW is massive, and it’s very easy to feel lost and isolated without supportive friends. For this reason, social organizations like Q++ are vital in providing students, especially ones that might struggle to immediately befriend a group of peers, a community. Additionally, Q++ has the potential to amplify and empower the voices of many LGBTQIA+ students within the Allen School itself. As a group, we can help advocate for meaningful changes that have the potential to better the student experience in the school for years to come. Together, there’s the added benefit that if someone is struggling, someone else in Q++ has likely faced a similar issue before. In this way, we can help guide each other through many of the common challenges we might experience at the UW.
Allen School: Why did you choose to major in computer science?
JS: I took my first computer science class my junior year of high school and from then on, I was hooked. When I was programming, however cliché it might sound, things seemed to click. I found that I wanted to continue coding after school, and sometimes looked forward to the homework assignments. Having never experienced this level of enthusiasm for any of my other studies, I thought that it had to mean something, so I applied to the Allen School. Now, having more knowledge about the field as a whole, I’ve decided to continue studying computer science because of how widespread software is. These days it seems like everything relies on computers and programming, meaning the opportunity is there to contribute to something that positively impacts the lives of people all around the world. I truly hope that by pursuing an education in computer science, I’ll get one of those opportunities to make a meaningful mark.
Allen School: What do you find most enjoyable about being an Allen School student?
JS: Personally, although there’s a whole lot that I’ve really enjoyed about being an Allen School student so far, I think that the overall community I’ve interacted with has really been the highlight. From Discord study groups to my grading parties with my fellow teaching assistants, everyone I’ve met has been open, welcoming, and more than willing to help you out if you get stuck on a tricky 311 proof. At the same time, the wealth of opportunities available at the Allen School makes being a student here all the better. As I mentioned above, I’m currently an undergraduate TA but I know people who are exploring research opportunities, participating in hack-a-thons, and contributing to codebases or working on personal projects. No matter what, there is always something to do at the Allen School if you take the time to look for it!
Another great thing about the school is that although I am studying computer science, the flexibility of the major has also allowed me to explore another of my passions: theatre. Even though there are many challenges with being a double degree student, I’m extremely grateful that I’ve been able to study two wildly different fields while at the UW.
Allen School: Who or what inspires you in the Allen School?
JS: This year especially, I think the resilience and perseverance of everyone in the Allen School is extremely inspiring. Sure, there have been issues and we’ve all faltered at some point, but the fact that everyone has worked through all the adversity that has surrounded this quarter is insanely impressive. Both the faculty and student body deserve some praise for how many adaptations have been made to make this quarter as normal as possible.
I also mentioned them earlier, but the TA community has been another force of inspiration throughout my time here in the Allen School. I’ve met countless peers who all share my passion for educating and spreading our knowledge of programming to fellow undergraduates, many of whom have no prior coding experience. The fact that so many within the Allen School are more than happy to share their understanding and unique perspective of computer science never ceases to amaze me and keeps me coming back quarter after quarter.
Thank you for your leadership and for supporting your fellow Allen School students in and out of the virtual classroom, Joe!
Leo Maddox Schneider: July 7, 2005 – January 12, 2019
As a student at Seattle’s Hamilton International Middle School, Leo Maddox Schneider demonstrated early mastery of mathematics and languages, was an avid gamer and athlete, and carved out a reputation as a budding conservationist. Enthusiastic about learning from an early age, Leo had already taken to heart his mother Sylvia Bolton’s advice to find something that he loved and was passionate about and to make that his profession. As she relayed to her son at the time, “it will bring fulfillment and a lot of happiness.”
What Leo loved was computer coding and Lego design; what he was passionate about were environmental causes. He might have pursued both at the University of Washington if not for the injuries he sustained in an automobile accident. Four and a half months later, on January 12, 2019, Leo passed away from those injuries and related complications. Nearly two years after that tragic loss, the foundation established by Leo’s mother to honor her son’s memory will give Allen School students the opportunity to fulfill their own dreams and carry on his legacy through the Leo Maddox Foundation Scholarship in Computer Science & Engineering.
”Leo loved computer science,” Bolton explained. “He and his friend Lennox shared a dream of attending a university that excelled in computer science so they could build their own company and make a difference in the world.”
Even at the tender age of 13, Leo was already well on his way toward making that difference. He forged enduring friendships with Lennox and Judson while playing Minecraft and Fortnite, which helped spark his interest in coding. He was already three years ahead of his grade level in mathematics and conversant in both Spanish and Bulgarian. His enthusiasm for the outdoors led Leo to champion environmental causes; he once convinced his mother to enter into one of their “non-negotiable” agreements permitting him to collect garbage for recycling. (Another of their non-negotiable agreements stipulated that he would eat his vegetables at dinner.) Leo was particularly passionate about the ocean, learning to swim with dolphins and developing a love of boat building craftsmanship inspired in part by his mother’s work as a luxury yacht designer.
“Everyone knew Leo as having a big, sweet soul and people just loved him. Losing him turned our world upside down into complete darkness,” recalled Bolton. “But we do not want the tragedy of Leo’s passing to define him. Leo was and will always be remembered as the smart, kind and compassionate kid who was gifted at math and science, loved the outdoors, and was a friend to many. With so much life ahead of him.”
To that end, Bolton established the Leo Maddox Foundation as a way to ensure that Leo’s legacy and aspirations for the future would live on in others. The Foundation supports a variety of initiatives designed to help promising young students with financial need to fully achieve their academic and creative potential, from assisting Rainier Scholars to go to college, to “Love, Leo” genius grants inspired by their namesake’s creative, can-do approach to solving problems he saw in the world. The new Leo Maddox Foundation Scholarship in Computer Science & Engineering will support Allen School undergraduate students in covering the cost of tuition and other educational expenses based on academic merit and financial need.
“We are heartbroken that Leo will never get the chance to apply to the Allen School and our hearts and prayers are with his family. We are deeply appreciative of the scholarship established by the Foundation in his name,” said professor Magdalena Balazinska, director of the Allen School. “This scholarship will touch many lives. It will promote the success of many talented students who need support to fulfill their dreams.”
In deference to her son’s twin loves, in addition to the Allen School scholarship Bolton also created the Leo Maddox Foundation Scholarship in Oceanography to support students in the School of Oceanography engaged in climate-related studies. The university’s preeminence in both disciplines and focus on student support convinced the Foundation to entrust it with Leo’s memory.
“As important as it is for the Leo Maddox Foundation to support young adults, it is equally important that we do so with the leaders in both fields,” said Vivian Ho, creator of the Leo Maddox Foundation. “In conducting our due diligence, it was clear that the University of Washington had a lot to offer in both areas of study and in shaping support for student scholarships. They created the perfect vehicles for our founder, Sylvia Bolton, to make the impactful difference she was seeking for Leo’s legacy.”
Last spring, Newport High School student Sophia Lin of Bellevue, Washington, was eager to start coding. Having applied to several summer coding programs, she was ready to learn. Unfortunately, as the pandemic spread, they were all canceled. Seeing her younger sister’s disappointment, Allen School senior Elizabeth Lin decided to be her coding teacher. With the help of her twin sister Christin, a senior in the University of Washington’s Department of Electrical & Computer Engineering, the two created a virtual summer coding program. Then they invited 300 kids from around the world to join them, free of charge
“As a teaching assistant in the Allen School’s Intro to Programming course, I had already taught hundreds of students how to code so I knew I could teach her, too,” said Elizabeth. “But knowing I had a busy summer with an internship at Microsoft as a software engineer, I decided to record the lessons. That way, I could teach others, too.”
Christin Lin
The two sisters reached out to their contacts from area high schools and advertised on Facebook to recruit 6th through 12th grade students interested in their new initiative, STEM League Developer Program, an offshoot from an outreach program the two started in high school.
“As big advocates for STEM education, we previously co-founded STEM League, an organization we started in high school with the mission to promote STEM opportunity and awareness to students in our local community,” said Christin. “Through STEM League, we hosted a series of STEM outreach programs at local libraries, facilitated Seattle STEM company tours, and volunteered for STEM events at local elementary schools. We wanted to continue that mission, therefore we launched the STEM League Developer Program.”
And within two weeks of recruiting students, the Lins had students from around the world signed up for the program with sponsorship from T-Mobile.
“The exponential interest in the Seattle area and around the world amazed us, but also brought some concerns,” said Elizabeth. “We worried about making sure every student had a quality experience. Learning to code can be daunting by yourself, and we wanted to ensure each student had support throughout the 8 weeks.”
From left to right: Elizabeth, Sophia and Christin Lin
The Lins once again returned to social media to recruit mentors — experienced high school and college programmers willing to work with the students. In about a month, the two screened and interviewed 50 mentors. By the end of June, they enrolled 300 students. Each mentor was assigned seven students to help, in various time zones across the world.
“Greg and I are thrilled when our interdisciplinary, hands-on entrepreneurship course contributes in at least some small way to giving students the inspiration and courage to head out and do remarkable things,” said Lazowska. “And what Elizabeth and her team have accomplished is truly remarkable.”
From Harry Potter-themed coding competitions, to a two-week program-wide hackathon where students created their own projects, to using Discord so everyone could stay in touch easily, the Lins and their collaborators found new and exciting ways to engage students and make coding more relevant to them.
“I thought the STEM League Developer Program was really fun, and it was definitely the highlight of my summer. I learned many important coding skills that I can now leverage for my future,” said one 10th grade participant. “Not only did I grow as a developer, I grew as a person, as I learned life skills like tenacity, motivation, and perseverance. The Discord community and the mentors were very supportive of everyone’s needs and helped create a positive learning environment where everyone could grow.”
Parents noticed how much their kids were enjoying the program. Several wrote in to thank the Lins and the mentors for teaching their kids during a particularly hard summer.
“My kids have been doing some other coding classes on a regular basis. But what they have learned at STEM League is of far better quality and better retained than any of those outside classes,” noted the parent of a student in Bellevue. “On top of it all they are having a lot of fun and they look forward to the next lesson. I don’t even have to keep nagging them to do it.”
According to Christin, the twins always wanted to host coding classes for students. They intend to continue the program next summer after they graduate.
Both sisters and five of the mentors in the program are alumni of the UW College of Engineering’s Washington State Academic RedShirt (STARS) program, a two-year specialized curriculum designed to build learning skills and strengthen academic preparation for core math and science prerequisites.
“This coding program idea sparked back in 2018, but we never found the time to do it,” she said. “However, this summer, we decided to launch it because we are both interested in entrepreneurship and experiencing something created by ourselves from scratch.”
UW, TCN and Althea members successfully deployed the first LTE in the Hilltop neighborhood of Tacoma last month.
While the internet is so critical for employment, education and communication, millions of Americans in rural and urban areas still do not have access to affordable connections. This lack of access further contributes to digital and economic inequality, especially during a pandemic when many schools and jobs have been moved online. A team of University of Washington researchers led by professor Kurtis Heimerl and Ph.D. student Esther Jang of the Allen School’s Information & Communications Technology for Development Lab are helping to address this problem right here in Washington.
With support from the Public Interest Technology University Network (PIT-UN), the group — which includes UW Tacoma urban studies professor Emma Slager and Jason Young, a senior research scientist in the iSchool — is deploying networks that will bring new, inexpensive, community-owned connectivity to marginalized communities in Seattle and Tacoma. The UW is one of 25 universities to receive a PIT-UN grant, which was created to fund critical research and build an inclusive career pipeline to advance the field of public interest technology.
Community cellular networks are owned and operated by the community they serve with the help of public and local organizations such as schools, nonprofits, community centers, makerspaces, libraries, small businesses, and tiny house villages. Leveraging their expertise in working with community cellular networks internationally, the team deployed a new network with a local connectivity non-profit, the Tacoma Community Network (TCN), to bring inexpensive, community-owned connectivity to the Hilltop community in the third largest city in Washington.
Bee Ivey, TCN’s volunteer executive director, said that as a nonprofit cooperative, TCN can focus on great speed and service without worrying about pleasing shareholders. To be sustainable, TCN needs about 20-25 members per “gateway,” which is an access point from which internet connectivity is distributed to the community. The partnership with the UW will empower TCN to connect more individuals at a lower price, which in turn allows them to get even more low income and extremely low income households online. Although the project is focused on making digital connections, it turns out that a personal touch was critical to making it work.
“We did a lot of canvassing prior to the pandemic, which allowed us to really connect with residents of Hilltop. One of the great things about being face-to-face is that you get to know people and hear their stories about how the internet affects their lives,” said Ivey. “It was truly eye-opening for us to meet so many people who didn’t have internet and had no way to access it, and definitely brought the statistics and research we’d seen to life.
Up to a quarter of all urban residents don’t have internet access, according to some studies, Ivey said, and it was made very clear the ways in which people are held back from living full, productive and satisfying lives when they lack internet access. Hearing these stories definitely strengthened the team’s resolve to continue the work connecting everyone to the internet. Now with the pandemic, they are focused on social media and networking, along with mailers to help reach more people who need internet access.
UW, TCN and Althea members built and deployed an LTE in Hilltop, in Tacoma
In Tacoma, one LTE network deployed in November contains eight households and is growing. Althea, a software company that makes mesh networking technology in which TCN’s routers use blockchain-based micro-payments to pay each other for traffic forwarding, is supporting the project. It has set up community wireless mesh networks all over the world, and is interested in integrating with LTE.
“Although we had a modest start, it represents a 30 percent adoption rate among the houses we were immediately able to reach,” said Ivey. “With the University of Washington’s help, we will be able to expand the number of households within our reach, as well as offer different types of internet connections — both typical wireless ISP equipment and LTE, the same data network used by cell phones. While there are many fantastic community-based internet networks out there, this particular type has never been deployed before in the United States to my knowledge, and it will make it far easier for individuals to access the internet.”
UW spun out the Local Connectivity Lab to deploy the LTE networking technology, powered by open-source software and operating in the Citizen’s Band Radio Service frequency spectrum, which is open enough to allow unlicensed devices to transmit in much of Seattle and Tacoma. This will allow the researchers to run open-sourced cellular networks in the U.S. on a small community scale.
“Cellular networks, with their higher-power access points, more favorable spectrum, and more efficient waveforms, have a much wider coverage area and user capacity than typical WiFi networks, and are also designed for user mobility like cell phones,” said Esther Jang, an Allen School graduate student leading the project. “Some initial line-of-sight link performance tests from our test deployment at UW yielded 60 megabits per second down and eight Mbps up with Consumer Premises Equipment, like a stationary user device with our SIM card at 1.3 miles away, from a backhaul connection around 150 Mbps.”
The team’s work is a part of the ICTD Lab’s goal to eventually create a cellular network that will allow people or organizations to deploy their own networks as easily as they do WiFi routers, where each network can come together to provide mutual roaming, which they call “cooperative cellular.” They are currently looking for non-profit organizations to help launch in Seattle, in addition to a King County Equity and Social Justice they recently received.
In addition to creating this open-sourced software and deploying it in communities that need it most, the group will also develop a STEM course called Community Networking. The course will give students an opportunity to explore the research, development and practice of access-related PIT and the partners and communities that are demonstrating an alternative viable path for a career in technology.
“Most of us take Internet access for granted to the point that, when the internet goes down, we struggle with continuing to get our work done,” said Allen School Director Magdalena Balazinska. “Yet some people, here in the United States, do not have such access. As computer scientists, we should always strive to solve important societal and world problems. I’m very excited about the way this project is using computer science to have a profound, positive impact on society.”
Pedro Domingos in the Paul G. Allen Center for Computer Science & Engineering. In his latest work, Domingos lifts the lid on the black box of deep learning. Dennis Wise/University of Washington
Deep learning has been immensely successful in recent years, spawning a lot of hope and generating a lot of hype, but no one has really understood why it works. The prevailing wisdom has been that deep learning is capable of discovering new representations of the data, rather than relying on hand-coded features like other learning algorithms do. But because deep networks are black boxes — what Allen School professor emeritus Pedro Domingos describes as “an opaque mess of connections and weights” — how that discovery actually happens is anyone’s guess.
Until now, that is. In a new paper posted on the preprint repository arXiv, Domingos gives us a peek inside that black box and reveals what is — and just as importantly, what isn’t — going on inside. Read on for a Q&A with Domingos on his latest findings, what they mean for our understanding of how deep learning actually works, and the implications for researchers’ quest for a “master algorithm” to unify all of machine learning.
You lifted the lid off the so-called black box of deep networks, and what did you find?
Pedro Domingos: In short, I found that deep networks are not as unintelligible as we thought, but neither are they as revolutionary as we thought. Deep networks are learned by the backpropagation algorithm, an efficient implementation for neural networks of the general gradient descent algorithm that repeatedly tweaks the network’s weights to make its output for each training input better match the true output. That process helps the model learn to label an image of a dog as a dog, and not as a cat or as a chair, for instance. This paper shows that all gradient descent does is memorize the training examples, and then make predictions about new examples by comparing them with the training ones. This is actually a very old and simple type of learning, called similarity-based learning, that goes back to the 1950s. It was a bit of a shock to discover that, more than half a century later, that’s all that is going on in deep learning!
Deep learning has been the subject of a lot of hype. How do you think your colleagues will respond to these findings?
PD: Critics of deep learning, of which there are many, may see these results as showing that deep learning has been greatly oversold. After all, what it does is, at heart, not very different from what 50-year-old algorithms do — and that’s hardly a recipe for solving AI! The whole idea that deep learning discovers new representations of the data, rather than relying on hand-coded features like previous methods, now looks somewhat questionable — even though it has been deep learning’s main selling point.
Conversely, some researchers and fans of deep learning may be reluctant to accept this result, or at least some of its consequences, because it goes against some of their deepest beliefs (no pun intended). But a theorem is a theorem. In any case, my goal was not to criticize deep learning, which I’ve been working in since before it became popular, but to understand it better. I think that, ultimately, this greater understanding will be very beneficial for both research and applications in this area. So my hope is that deep learning fans will embrace these results.
So it’s a good news/bad news scenario for the field?
PD: That’s right. In “The Master Algorithm,” I explain that when a new technology is as pervasive and game-changing as machine learning has become, it’s not wise to let it remain a black box. Whether you’re a consumer influenced by recommendation algorithms on Amazon, or a computer scientist building the latest machine learning model, you can’t control what you don’t understand. Knowing how deep networks learn gives us that greater measure of control.
So, the good news is that it is now going to be much easier for us to understand what a deep network is doing. Among other things, the fact that deep networks are just similarity-based algorithms finally helps to explain their brittleness, whereby changing an example just slightly can cause the network to make absurd predictions. Up until now, it has puzzled us why a minor tweak would, for example, lead a deep network to suddenly start labeling a car as an ostrich. If you’re training a model for a self-driving car, you probably don’t want to hit either, but for multiple reasons — not least, the predictability of what an oncoming car might do compared to an oncoming ostrich — I would like the vehicle I’m riding in to be able to tell the difference.
But these findings could be considered bad news in the sense that it’s clear there is not much representation learning going on inside these networks, and certainly not as much as we hoped or even assumed. How to do that remains a largely unsolved problem for our field.
If they are essentially doing 1950s-style learning, why would we continue to use deep networks?
PD: Compared to previous similarity-based algorithms such as kernel machines, which were the dominant approach prior to the emergence of deep learning, deep networks have a number of important advantages.
One is that they allow incorporating bits of knowledge of the target function into the similarity measure — the kernel — via the network architecture. This is advantageous because the more knowledge you incorporate, the faster and better you can learn. This is a consequence of what we call the “no free lunch” theorem in machine learning: if you have no a priori knowledge, you can’t learn anything from data besides memorizing it. For example, convolutional neural networks, which launched the deep learning revolution by achieving unprecedented accuracy on image recognition problems, differ from “plain vanilla” neural networks in that they incorporate the knowledge that objects are the same no matter where in the image they appear. This is how humans learn, by building on the knowledge they already have. If you know how to read, then you can learn about science much faster by reading textbooks than by rediscovering physics and biology from scratch.
Another advantage to deep networks is that they can bring distant examples together into the same region, which makes learning more complex functions easier. And through superposition, they’re much more efficient at storing and matching examples than other similarity-based approaches.
Can you describe superposition for those of us who are not machine learning experts?
PD: Yes, but we’ll have to do some math. The weights produced by backpropagation contain a superposition of the training examples. That is, the examples are mapped into the space of variations of the function being learned and then added up. As a simple analogy, if you want to compute 3 x 5 + 3 x 7 + 3 x 9, it would be more efficient to instead compute 3 x ( 5 + 7 + 9) = 3 x 21. The 5, 7 and 9 are now “superposed” in the 21, but the result is still the same as if you separately multiplied each by 3 and then added the results.
The practical result is that deep networks are able to speed up learning and inference, making them more efficient, while reducing the amount of computer memory needed to store the examples. For instance, if you have a million images, each with a million pixels, you would need on the order of terabytes to store them. But with superposition, you only need an amount of storage on the order of the number of weights in the network, which is typically much smaller. And then, if you want to predict what a new image contains, such as a cat, you need to cycle through all of those training images and compare them with the new one. That can take a long time. With superposition, you just have to pass the image through the network once. That takes much less time to execute. It’s the same with answering questions based on text; without superposition, you’d have to store and look through the corpus, instead of a compact summary of it.
So your findings will help to improve deep learning models?
PD: That’s the idea. Now that we understand what is happening when the aforementioned car suddenly becomes an ostrich, we should be able to account for that brittleness in the models. If we think of a learned model as a piece of cheese and the failure regions as holes in that cheese, we now understand better where those holes are, and what their shape and size is. Using this knowledge, we can actively figure out where we need new data or adjustments to the model to fix the holes. We should also improve our ability to defend against attacks that cause deep networks to misclassify images by tweaking some pixels such that they cause the network to fall into one of those holes. An example would be attempts to fool self-driving cars into misrecognizing traffic signs.
What are the implications of your latest results in the search for the master algorithm?
PD: These findings represent a big step forward in unifying the five major machine learning paradigms I described in my book, which is our best hope for arriving at that universal learner, what I call the “master algorithm.” We now know that all learning algorithms based on gradient descent — including but not limited to deep networks — are similarity-based learners. This fact serves to unify three of the five paradigms: neural, probabilistic, and similarity-based learning. Tantalizingly, it may also be extensible to the remaining two, symbolic and genetic learning.
Given your findings, what’s next for deep learning? Where does the field go from here?
PD: I think deep learning researchers have become too reliant on backpropagation as the near-universal learning algorithm. Now that we know how limited backprop is in terms of the representations it can discover, we need to look for better learning algorithms! I’ve done some work in this direction, using combinatorial optimization to learn deep networks. We can also take inspiration from other fields, such as neuroscience, psychology, and evolutionary biology. Or, if we decide that representation learning is not so important after all — which would be a 180-degree change — we can look for other algorithms that can form superpositions of the examples and that are compact and generalize well.
The American Association for the Advancement of Science, the world’s largest general scientific society, has named Allen School professor emeritus Pedro Domingos and professor Daniel Weld among its class of 2020 AAAS Fellows honoring members whose scientifically or socially distinguished efforts have advanced science or its applications. Both Domingos and Weld were elected Fellows in the organization’s Information, Computing, and Communication section for their significant impact in artificial intelligence and machine learning research.
Domingos was honored by the AAAS for wide-ranging contributions in AI spanning more than two decades and 200 technical publications aimed at making it easier for machines to discover new knowledge, learn from experience, and extract meaning from data with little or no help from people. Prominent among these, to his AAAS peers, was his introduction of Markov logic networks unifying logical and probabilistic reasoning. He and collaborator Matthew Richardson (Ph.D., ‘04) were, in fact, the first to coin the term Markov logic networks (MLN) when they presented their simple yet efficient approach that combined first-order logic and probabilistic graphical models to support inference learning.
Domingos’ work has resulted in several other firsts that represented significant leaps forward for the field. He again applied Markov logic to good effect to produce the first unsupervised approach to semantic parsing — a key method by which machines extract knowledge from text and speech and a foundation of machine learning and natural language processing — in collaboration with then-student Hoifung Poon (Ph.D., ‘11). Later, Domingos worked with graduate student Austin Webb (M.S., ‘13) on Tractable Markov Logic (TML), the first non-trivially tractable first-order probabilistic language that suggested efficient first-order probabilistic inference could be feasible on a larger scale.
Domingos also helped launch a new branch of AI research focused on adversarial learning through his work with a team of students on the first algorithm to automate the process of adversarial classification, which enabled data mining systems to adapt in the face of evolving adversarial attacks in a rapid and cost-effective way. Among his other contributions was the Very Fast Decision Tree learner (VFDT) for mining high-speed data streams, which retained its status as the fastest such tool available for 15 years after Domingos and Geoff Hulten (Ph.D., ‘05) first introduced it.
In line with the AAAS’ mission to engage the public in science, in 2015 Domingos published The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World. Geared to the expert and layperson alike, the book offers a comprehensive exploration of how machine learning technologies influence nearly every aspect of people’s lives — from what ads and social posts they see online, to what route their navigation system dictates for their commute, to what movie a streaming service suggests they should watch next. It also serves as a primer on the various schools of thought, or “tribes,” in the machine learning field that are on a quest to find the master algorithm capable of deriving all the world’s knowledge from data.
Prior to this latest honor, Domingos was elected a Fellow of the Association for the Advancement of Artificial Intelligence (AAAI) and earned two of the highest accolades in data science and AI: the SIGKDD Innovation Award from the Association of Computing Machinery’s Special Interest Group on Knowledge Discovery and Data Mining, and the IJCAI John McCarthy Award from the International Joint Conference on Artificial Intelligence.
AAAS recognized Weld for distinguished contributions in automated planning, software agents, crowdsourcing, and internet information extraction during a research career that spans more than 30 years. As leader of the UW’s Lab for Human-AI Interaction, Weld seeks to combine human and machine intelligence to accomplish more than either could on their own. To that end, he and his team focus on explainable machine learning, intelligible and trustworthy AI, and human-AI team architectures to enable people to better understand and control AI-driven tools, assistants, and systems.
Weld has focused much of his career on advanced intelligent user interfaces for enabling more seamless human-machine interaction. Prominent among these is SUPPLE, a system he developed with Kryzstof Gajos (Ph.D., ‘08) that dynamically and optimally renders user interfaces based on device characteristics and usage patterns while minimizing user effort. Recognizing the potential for that work to improve the accessibility of online tools for people with disabilities, the duo subsequently teamed up with UW Information School professor and Allen School adjunct professor Jacob Wobbrock to extend SUPPLE’s customization to account for a user’s physical capabilities as well.
Another barrier that Weld has sought to overcome is the amount of human effort required to organize and maintain the very large datasets that power AI applications. To expedite the process, researchers turned to crowdsourcing, but the sheer size and ever-changing nature of the datasets still made it labor-intensive. Weld, along with Jonathan Bragg (Ph.D., ‘18) and affiliate faculty member Mausam (Ph.D., ‘07), created Deluge to optimize the process of multi-label classification that significantly reduced the amount of labor required compared to the previous state of the art without sacrificing quality. Quality control is a major theme of Weld’s work in this area, which has yielded new tools such as Sprout for improving task design, MicroTalk and Cicero for augmenting decision-making, and Gated Instruction for more accurate relation extraction.
In addition to his technical contributions, AAAS also cited Weld’s impact via the commercialization of new AI technologies. During his tenure on the UW faculty, he co-founded multiple venture-backed companies based on his research: Netbot Inc., creator of the first online comparison shopping engine that was acquired by Excite; AdRelevance, an early provider of tools for monitoring online advertising data that was acquired by Nielsen Netratings; and Nimble Technology, a provider of business intelligence software that was acquired by Actuate. Weld has since gone from founder to funder as a venture partner and member of the Technology Advisory Board at Madrona Venture Group.
Weld, who holds the Thomas J. Cable/WRF Professorship, presently splits his time between the Allen School, Madrona, and the Allen Institute for Artificial Intelligence (AI2), where he directs the Semantic Scholar research group focused on the development of AI-powered research tools to help scientists overcome information overload and extract useful knowledge from the vast and ever-growing trove of scholarly literature. Prior to this latest recognition by AAAS, Weld was elected a Fellow of both the AAAI and the ACM. He is the author of roughly 200 technical papers and two books on AI on the theories of comparative analysis and planning-based information agents, respectively.
Domingos and Weld are among four UW faculty members elected as AAAS Fellows this year. They are joined by Eberhard Fetz, a professor in the Department of Physiology & Biophysics and DXARTS who was honored in the Neuroscience section for his contributions to understanding the role of the cerebral cortex in controlling ocular and forelimb movements as well as motor circuit plasticity, and Daniel Raftery, a professor in UW Medicine’s Department of Anesthesiology and Pain Medicine who was honored in the Chemistry section for his contributions in the fields of metabolomics and nuclear magnetic resonance, including advanced analytical methods for biomarker discovery and cancer diagnosis.
The robust impact that the Allen School and the University of Washington have in contributing to accessible technology was recognized at the 22nd International ACM SIGACCESS Conference on Computer and Accessibility (ASSETS 2020) held virtually last month. Researchers from the Allen School and the UW contributed to the Best Student Paper, Best Artifact and the Best Paper.
A team led by UW Human Centered Design & Engineering alumna and Carnegie Mellon University postdoc Cynthia Bennett earned the Best Student Paper Award for Living Disability Theory: Reflections on Access, Research and Design. The paper was co-written by lead author Megan Hoffman, a Ph.D. student at CMU, along with Allen School professor Jennifer Mankoff and City University of New York professor Devva Kasnitz. The paper emphasizes the importance of integrating disability studies perspectives and disabled people into accessibility research.
Top left to right: Mankoff, Froehlich, Jain Bottom left to right: Patel, Ngo, Nguyen
In the paper, the authors correlate personal experiences with theoretical experiences. They found that while accessibility research tends to focus on creating technology related to impairment, without including disability studies — which seeks to understand disability and advocate against ableist systems — accessibility research isn’t as inclusive as its intended purpose. From their research and personal experiences, the authors exemplify how disability is often mired in ableism and oversimplified. They urge disability researchers to commit to recognizing and repairing ableism; study disability beyond diagnosis; incorporate a disability studies perspective that includes disabled voices; and incorporate reflexive, interpretivist study as a regular and essential practice.
“It was so inspiring to learn from and be part of the team writing this paper,” said Mankoff, who leads the Allen School’s Make4All group. “More than anything it showed me that the next generation of scholars are already leading the way in defining what matters in our scholarship.”
Members of the Allen School also contributed to the Best Artifact: SoundWatch, a smartwatch app for d/Deaf and hard-of-hearing people who want to be aware of nearby sounds. The creators are Allen School Ph.D. student and lead author Dhruv Jain, Pratyush Patel and professor Jon Froehlich; undergraduates Hung Ngo and Khoa Nguyen; HCDE professor and Allen School adjunct professor Leah Findlater; Ph.D. student Steven Goodman; and research scientist Rachel Grossman-Kahn.
SoundWatch is an app for Android smartwatches that uses machine learning to alert users of sounds like nearby fire alarms and beeping microwaves, making their environment more accessible. Soundwatch identifies the sound and alerts the user with a friendly buzz along with information about the sound on the screen of the watch.
“This technology provides people with a way to experience sounds that require an action — such as getting food from the microwave when it beeps. But these devices can also enhance people’s experiences and help them feel more connected to the world,” said Jain in a recent UW News release. “I use the watch prototype to notice birds chirping and waterfall sounds when I am hiking. It makes me feel present in nature. My hope is that other d/Deaf and hard-of-hearing people who are interested in sounds will also find SoundWatch helpful.”
“The University of Washington has been a leader in accessible technology research, design, engineering, and evaluation for years,” said iSchool professor and Allen School adjunct professor Jacob O. Wobbrock, who, along with Mankoff, serves as founding co-director of the UW’s Center for Research and Education on Accessible Technology and Experiences (CREATE).”This latest round of awards from ACM ASSETS is further testament to the great work being done at the UW. Now, with the recent launch of CREATE, our award-winning faculty and students are brought together like never before, and we are already seeing the great things that come of it.”
Congratulations to all of the ASSETS 2020 award recipients!
Robots have traditionally been deployed for dull, dirty or dangerous tasks. What if robots instead could be used to support the sophisticated and iterative work of domain experts such as chemical engineers or synthetic biologists?
A University of Washington research project led by Allen School adjunct faculty member and Human-Centered Design and Engineering professor Nadya Peek and Allen School and Electrical and Computer Engineering professor Josh Smith, “NRI: FND: Multi-Manipulator Extensible Robotic Platforms,” received a $700,000 grant from the National Science Foundation’s (NSF) National Robotics Initiative 2.0: Ubiquitous Collaborative Robots (NRI-2.0) program.
“The tools we propose to develop include a family of open-source, replicable, extensible, parametrically-defined co-bots that will enable experts to iteratively develop automated processes and experiments,” Peek said. “This grant will help us develop hardware and software for authoring, running, and verifying automated workflows.”
Smith’s lab has developed an ultrasonic manipulator that allows a robot to pick up small objects without touching them. The grant will allow the researchers to combine this new ultrasonic manipulator with Peek’s open source multi-tool motion platforms, including Jubilee.
“Non-contact manipulation can allow robots to pick up small objects and powders, which is currently challenging for robots,” Smith said. “Non-contact manipulation can also help maintain sterility, which could be useful in surgical settings, and any time we are concerned about spreading pathogens.”
The integrated robotic system will allow end-users to develop automated workflows for domain specific tasks. The researchers are designing their system to be customizable and extensible. In particular, the robotic systems they develop are fabricatable, meaning that they can be made with easily sourced parts or parts made using low cost digital fabrication tools such as 3D printers. This means that even when the domain experts create highly sophisticated interactive and automated workflows, their experimental setups can easily be reproduced by other scientists.
