Allen School News

Fascinated by the inner workings of machine learning models for data-driven decision-making, Allen School professor Simon Shaolei Du constructs their theoretical foundations to better understand what makes them tick and then designs algorithms that translate theory into practice. Du’s faculty colleague Adriana Schulz, meanwhile, has clocked how to make the act of making more accessible and sustainable through novel techniques in computer-aided design and manufacturing, drawing upon advances in machine learning, fabrication, programming languages and more. 

Those efforts received a boost from the Alfred P. Sloan Foundation earlier this year, when Schulz and Du were recognized among the 2024 class of Sloan Research Fellows representing the next generation of scientific leaders.

Simon Shaolei Du: Unlocking the mysteries of machine learning

Deep learning. Reinforcement learning. Representation learning. Recent breakthroughs in the training of large-scale machine learning models are transforming data-driven decision-making across a variety of domains and fueling developments ranging from self-driving vehicles to ChatGPT. But while we know that such models work, we don’t really know why.

“We still don’t have a good understanding of why these paradigms are so powerful,” Du explained in a UW News release. “My research aims to open the black box.”

Already, Du has been able to poke several holes in said box by demystifying several principles underlying the success of such models. For example, Du offered the first proof for how gradient descent optimizes the training of over-parameterized deep neural networks — so-called because the number of parameters significantly exceeds the minimum required relative to the size of the training dataset. Du and his co-authors showed that, with sufficient over-parameterization, gradient descent could find the global minima to achieve zero training loss even though the objective function is non-convex and non-smooth. Du was also able to explain how these models generalize so well despite their enormous size by proving a fundamental connection between deep neural network learning and kernel learning. 

Another connection Du has investigated is that between representation learning and recent advances in computer vision and natural language processing. Representation learning bypasses the need to train on each new task from scratch by drawing upon the commonalities underlying different but related tasks. Du was keen to understand how using large-scale but low-quality data to pre-train foundation models in the aforementioned domains effectively improves their performance on downstream tasks for which data is scarce — a condition known as few-shot learning. He and his collaborators developed a novel theoretical explanation for this phenomenon by proving that a good representation combined with a diversity of source training data are both necessary and sufficient for few-shot learning on a target task. Following this discovery, Du contributed to the first active learning algorithm for selecting pre-training data from the source task based on their relevance to the target task to make representation learning more efficient.

From representation to reinforcement: When it comes to modeling problems in data-driven decision-making, the latter is the gold standard. And the standard wisdom is that long planning horizons and large state spaces are why it is so difficult — or at least it was. Du and his collaborators turned the first assumption on its head by showing that sample complexity in reinforcement learning is not dependent upon whether the planning horizon is long or short. Du further challenged prevailing wisdom by demonstrating that a good representation of the optimal value function — which was presumed to address the state space problem — is not sufficient to ensure sample-efficient reinforcement learning across states. 

“My goal is to design machine learning tools that are theoretically principled, resource-efficient and broadly accessible to practitioners across a variety of domains,” said Du. “This will also help us to ensure they are aligned with human values, because it is apparent that these models are going to play an increasingly important role in our society.”

Adriana Schulz: Making a mark by remaking manufacturing-oriented design

AI’s influence on design is already being felt in a variety of sectors. But despite its promise to enhance quality and productivity, its application to design for manufacturing has lagged. So, too, has the software side of the personalized manufacturing revolution, which has failed to keep pace with hardware advances in 3D-printing, machine knitting, robotics and more. This is where Schulz aims to make her mark.

“Design for manufacturing is where ideas are transformed into products that influence our daily lives,” Schulz said. “We have the potential to redefine how we ideate, prototype and produce almost everything.”

To realize this potential, Schulz develops computer-aided design tools for manufacturing that are grounded in the fundamentals of geometric data processing and physics-based modeling and also draw from domains such as machine learning and programming languages. The goal is to empower users of varying skill levels and backgrounds to flex their creativity while optimizing their designs for functionality and production. 

One strategy is to treat design and fabrication as programs — that is, a set of physical instructions — and leverage formal reasoning and domain-specific languages to enable users to adjust plans on the fly based on their specific goals and constraints. Schulz and her collaborators took this approach with Carpentry Compiler, a tool for exploring tradeoffs between production time, cost of materials and other factors of their design before generating fabrication plans. She subsequently parlayed advances in program synthesis into a new tool for efficiently optimizing plans for both design and fabrication at the same time. Leveraging a technique called equivalence graphs, or e-graphs, Schulz and her team took advantage of inherent redundancies across design variations and fabrication alternatives to eliminate the need to recompute the fabrication cost from scratch with every design change. In a series of experiments, the new framework was shown to reduce project costs by as much as 60%.

Rising capabilities in AI have also given rise to a new field in computer science known as neurosymbolic reasoning, a hybrid approach to representing visual and other types of data that combines techniques from machine learning and symbolic program analysis. Schulz leveraged this emerging paradigm to make it easier for users of parametric CAD models for manufacturing to explore and manipulate variations of their designs while automatically retaining essential structural constraints. Typically, CAD users who wish to engage in such exploration have to go to the time and trouble of modifying multiple parameters simultaneously and then sifting through a slew of irrelevant outcomes to identify the meaningful ones. Schulz and her team streamlined the process by employing large language and image models to infer the space of meaningful variations of a shape, and then applying symbolic program analysis to identify common constraints across designs. Their system, ReparamCAD, offers a more intuitive, efficient and interactive approach to conventional CAD programs.

In addition to introducing more flexible design processes, Schulz has also contributed to more flexibility on the factory floor. Many assembly lines rely on robots that are task-specific, making it complex and costly to pivot the line to new tasks. Schulz and her colleagues sidestepped this problem by enabling the creation of 3D-printable passive grippers that can be swapped out at the end of a robotic arm to handle a variety of objects — including irregular shapes that would be a challenge for conventional grippers to manipulate. She and her team developed an algorithm that, when fed a 3D model of an object and its orientation, co-optimizes a gripper design and lift trajectory that will enable the robot to successfully pick up the item.

Whether it’s repurposed robots or software that minimizes material waste, Schulz’s past work offers a glimpse into manufacturing’s future — one that she hopes will be friendlier not just to people, but also to the planet.

“Moving forward, I plan to expand my efforts on sustainable design, exploring innovative design solutions that prioritize reusability and recyclability to foster circular ecosystems,” she told UW News.

Two other researchers with Allen School connections were among 22 computer scientists across North America to have been recognized among the 2024 class of Sloan Research Fellows. Justine Sherry (B.S., ‘10) is a professor at Carnegie Mellon University, where she leads research to modernize hardware and software for implementing middleboxes to make the internet more reliable, efficient, secure and equitable for users. Former visiting student Arvind Satyanarayan, who earned his Ph.D. from Stanford University while working with Allen School professor Jeffrey Heer in the UW Interactive Data Lab, is a professor at MIT, where he leads the MIT Visualization Group using interactive data visualization to explore intelligence augmentation that will amplify creativity and cognition while respecting human agency.

In addition, a third UW faculty member, chemistry professor Alexandra Velian, earned a Sloan Research Fellowship for her work on new materials to advance decarbonization, clean energy and quantum information technologies. 

For more on the 2024 Sloan Research Fellows, see the Sloan Foundation’s announcement and a related story by UW News.

Read more →

For as long as she can remember, Allen School professor Su-In Lee wanted to be a scientist and professor when she grew up. Her father, who majored in marine shipbuilding engineering, would sit her down at their home in Korea with a pencil and paper to teach her math. Those childhood lessons instilled in Lee not just a love of the subject matter but also of teaching; her first pupil was her younger brother, with whom she shared what she had learned about arithmetic and geometry under her father’s tutelage. 

Fast forward a few decades, and Lee is putting those lessons to good use in training the next generation of scientists and engineers while advancing explainable artificial intelligence for biomedical applications. She is also adding up the accolades in response to her work. In February, the International Society for Computational Biology recognized Lee with its 2024 ISCB Innovator Award, given to a mid-career scientist who has consistently made outstanding contributions to the field of computational biology and continues to forge new directions; last month, the American Institute for Medical and Biological Engineering inducted her into the AIMBE College of Fellows — putting her among the top 2% of medical and biological engineers. Yesterday, the Ho-Am Foundation announced Lee as the 2024 Samsung Ho-Am Prize Laureate in Engineering for her pioneering contributions to the field of explainable AI.

As the saying goes, good things come in threes.

“This is an incredible honor for me, and I’m deeply grateful for the recognition,” said Lee, who holds the Paul G. Allen Professorship in Computer Science & Engineering and directs the AI for bioMedical Sciences (AIMS) Lab at the University of Washington. “There are so many deserving researchers, I am humbled to have been chosen. One of the most rewarding aspects of my role as a faculty member and scientist is serving as an inspiration for young people. As AI continues to transform science and society, I hope this inspires others to tackle important challenges to improve health for everyone.”

The Ho-Am Prize, which is often referred to as the “Korean Nobel Prize,” honors people of Korean heritage who have made significant contributions in academics, the arts and community service or to the welfare of humanity through their professional achievements. Previous laureates include Fields Medal-winning mathematician June Huh and Academy Award winning director Joon-ho Bong of “Parasite” fame. In addition to breaking new ground through her work, Lee has broken the glass ceiling: She is the first woman to receive the engineering prize in the award’s 34-year history and, still in her 40s, one of the youngest recipients in that category — a testament to the outsized impact she has made so early in her career.

From 1-2-3 to A-B-C

As a child, she may have learned to love her 1-2-3s; as a researcher, Lee became more concerned with her A-B-Cs: AI, biology and clinical medicine.

“The future of medicine hinges on the convergence of these disciplines,” she said. “As electronic health records become more prevalent, so too will omic data, where AI will play a pivotal role.”

Before her arrival at Stanford to pursue her Ph.D., the “C” could have stood for “cognition,” which marked her first foray into researching AI models like deep neural networks. For her undergraduate thesis, she developed a DNN for hand-written digit recognition that won the 2000 Samsung Humantech Paper Award. In the Stanford AI Lab, Lee shifted away from cognition and toward computational molecular biology, enticed by the prospect of identifying cures for diseases such as Alzheimer’s. She continued to be captivated by such questions after joining the Allen School faculty in 2010.

Six years after her arrival, Lee’s research took an unexpected — but welcome — turn when Gabriel Erion Barner knocked on her office door. Erion Barner was a student in the UW’s Medical Scientist Training Program, or MSTP, and he had a proposition.

“MSTP students combine a medical degree with a Ph.D. in a complementary field, and they are amazing,” said Lee. “Gabe was excited about AI’s potential in medicine, so he decided he wanted to do a Ph.D. in Computer Science & Engineering working with me. There was just one problem: our Ph.D. program didn’t have a process to accommodate students like Gabe. So, we created one.”

Erion Barner formally enrolled the following year and became the first MSTP student to graduate with an Allen School degree, in the spring of 2021. But he wouldn’t be the last. Joseph Janizek (Ph.D., ‘22) and current M.D. student Alex DeGrave subsequently sought out Lee as an advisor. Erion Barner has since moved on to Harvard Medical School to complete his medical residency, while Janizek is about to do the same at Lee’s alma mater, Stanford.

Meanwhile, Lee’s own move into the clinical medicine aspect of the A-B-Cs was complete.

Getting into SHAP

Beyond the A-B-Cs, Lee subscribes to a philosophy she likens to latent factor theory. A term borrowed from machine learning, latent factor theory posits that there are underlying — and unobserved — factors that impact upon the observable ones. Lee applies this theory when selecting the research questions in which she will invest her time. It’s part of her quest to identify that underlying factor which will transcend multiple problems, domains and disciplines.

So, when researchers began applying AI to medicine, Lee was less interested in making the models’ predictions more accurate in favor of understanding why they made the predictions they did in the first place.

“I just didn’t want to do it,” she said of pursuing the accuracy angle. “Of course we need the models to be accurate, but why was a certain prediction made? I realized that addressing the black box of AI — those latent factors — would be helpful for clinical decision-making and for clinicians’ perceptions of whether they could trust the model or not.” This principle, she noted, extends beyond medical contexts to areas like finance.

Lee discovered that the questions she raised around transparency and interpretability were the same questions circulating in the medical community. “They don’t just want to be warned,” she said. “They want to know the reasons behind the warning.”

In 2018, Lee and her team began to shine a lot on the models’ reasoning. Their first clinical paper, appearing on the cover of Nature Biomedical Engineering, described a novel framework for not only predicting but also providing real-time explanations for a patient’s risk of developing hypoxemia during surgery. The framework, called Prescience, relied on SHAP values — short for SHapley Additive exPlanations — which applied a game theoretic approach to explain the weighted outputs of a model. The paper, which is broadly applicable to many domains, has garnered more than 1,300 citations. In follow-up work, Lee and her team unveiled CoAI, or Cost-Aware Artificial Intelligence, which applied Shapley values to prioritize which patient risk factors to evaluate in emergency or critical care scenarios given a budget of time, resources or both.

AI under the microscope

Lee and her collaborators subsequently shifted gears in the direction of developing fundamental AI principles and techniques that could transcend any single clinical or research question. Returning to her molecular biology roots, Lee was curious about how she could use explainable AI to solve common problems in analyzing single-cell datasets to understand the mechanisms and treatment of disease. But to do that, researchers would need to run experiments in which they disentangle variations in the target cells from those in the control dataset to identify which factors are relevant and which merely confound the results. To that end, Lee and her colleagues developed ContrastiveVI, a deep learning framework for applying a technique known as contrastive analysis to single-cell datasets. The team published their framework in Nature Methods.

“By addressing contrastive scientific questions, we can help solve many problems,” Lee explained. “Our methods enable us to handle these nuanced datasets effectively.”

Up to that point, the utility of CA in relation to single cell data was limited; for once, latent factors — in this case, the latent variables typically used to model all variations in the data — worked against the insights Lee sought. ContrastiveVI solves this problem by separating those latent variables, which are shared across both the target and control datasets, from the salient variables exclusive to the target cells. This enables comparisons of, for example, the differences in gene expression between diseased and healthy tissue, the body’s response to pathogens or drugs, or CRISPR-edited versus unedited genomes.

Lee and her colleagues also decided to put medical AI models themselves under the microscope, applying more scrutiny to their predictions by developing techniques to audit their performance in domains ranging from dermatology to radiology. As they discovered, even when a model’s predictions are accurate, they should be treated with a healthy dose of skepticism.

“I’m particularly drawn to this direction because it underscores the importance of understanding AI models’ reasoning processes before blindly using them — a principle that extends across disciplines, from single-cell foundation models, to drug discovery, to clinical trial identification,” she said.

As one example, Lee and her team revealed how models that analyze chest x-rays to predict whether a patient has COVID-19 tend to rely on so-called shortcut learning, which leads them to base their predictions on spurious factors rather than genuine medical pathology. The journal Nature highlighted their work the following year. More recently, Lee co-authored a paper in Nature Biomedical Engineering that leveraged generative AI to audit medical-image classifiers used for predicting melanoma, finding that they rely on a mix of clinically significant factors and spurious associations. The team’s method, which favored the use of counterfactual images over conventional saliency maps to make the image classifiers’ predictions medically understandable, could be extended to other domains such as radiology and ophthalmology. Lee recently published her futuristic perspectives on the clinical potential of counterfactual AI in The Lancet.

Going forward, Lee is “really excited” about tackling the black-box nature of medical-image classifiers by automatically annotating semantically meaningful concepts using image-text foundation models. In fact, she has a paper on this very topic that is slated to be published in Nature Medicine this spring.

“All research topics hold equal fascination for me, but if I had to choose one, our AI model auditing framework stands out,” Lee said. “This unique approach can be used to uncover flaws in the reasoning process of AI models, which could solve a lot of society’s concerns about AI. We humans have a tendency to fear the unknown, but our work has demonstrated that AI is knowable.”

Live long and prosper

From life-and-death decisionmaking in the ER, to a more nuanced approach to analyzing life and death: One of the topics Lee has investigated recently concerns applying AI techniques to massive amounts of clinical data to understand people’s biological ages. The ENABL Age framework — which stands for ExplaiNAble BioLogical Age — applied explainable AI techniques to all-cause and cause-specific mortality to predict individuals’ biological ages and identify the underlying risk factors that contributed to those predictions to potentially inform clinical decision-making. The paper was featured on the cover of Lancet Healthy Longevity last December.

Lee hopes to build on this work to uncover the drivers of aging as well as the underlying mechanisms of rejuvenation — a topic she looks forward to exploring with her peers at a workshop she is co-leading in May. She is also keen to continue applying AI insights to identify therapeutic targets for Alzheimer’s disease, which is one of the 10 deadliest diseases in the U.S., as well as other neurodegenerative conditions. Her prior work on this topic was published in Nature Communications in 2021 and Genome Biology in 2023.

Even with AI’s flaws — some of which she herself has uncovered — Lee believes that it will prove to be a net benefit to society.

“Like other technologies, AI carries risks,” Lee acknowledged. “But those risks can be mitigated through the use of complementary technologies such as explainable AI that allow us to interpret complex AI models to promote transparency, accountability, and ultimately, trust.”

If Lee’s father, who passed away in 2013, could see her now, he would no doubt be impressed with how those early math lessons added up.

Read more →

With large language models dominating the discourse these days, artificial intelligence researchers find themselves increasingly in the limelight. But while LLMs continue to grow in size — and capture a growing share of the public’s imagination — their utility could be limited by their voracious appetite for compute resources and power. 

This is where the systems researchers have an opportunity to shine. And it so happens that one of the brightest sparks working at the intersection of AI and systems can be found right here at the University of Washington. 

Zihao Ye, a fourth-year Ph.D. student in the Allen School, builds serving systems for foundation models and sparse computation to improve the efficiency and enhance the programmability of emerging architectures such as graph networks and the aforementioned LLMs. To support his efforts, NVIDIA recently selected Ye as one of 10 recipients of the company’s highly competitive Graduate Research Fellowship. The honorees are described by NVIDIA Chief Scientist Bill Dally as “among the most talented graduate students in the world.”

Ye applies his talents to the development of techniques that enable machine learning systems with large and sparse tensors — and their large workloads — to run more efficiently in resource-constrained contexts such as smartphones and web browsers. To that end, he teamed up with professor Luis Ceze and alum Tianqi Chen (Ph.D., ’19), now a faculty member at Carnegie Mellon University and co-founder alongside Ceze of Allen School spinout OctoAI, in the Allen School’s interdisciplinary SAMPL group.

“Zihao is a deep thinker who is diligent about background research and extremely skilled in systems building. That is a powerful combination in a systems researcher,” said Ceze, who holds the Edward D. Lazowska Professorship in Computer Science & Engineering at the Allen School and also serves as CEO of OctoAI. “He also has a good eye for research problems and is a fantastic colleague and teammate.”

Ye’s eye for research problems led him to pursue what Ceze termed a “very elegant idea” for overcoming the so-called hardware lottery when programming neural networks to run on modern GPUs. One of the main obstacles is that neural networks, such as those used in graph analytics, are sparse tensor applications, whereas modern hardware is designed primarily for dense tensor operations. To solve the problem, Ye and his colleagues created SparseTIR, a composable programming abstraction that supports efficient sparse model optimization and compilation. SparseTIR decomposes a sparse matrix into multiple sub-matrices with homogeneous sparsity patterns to enable more hardware-friendly storage, while offloading the associated computation to different compute units within GPUs to optimize runtime performance. The team layered their approach onto Apache TVM, an open-source framework that supports the deployment of machine learning workloads on any hardware backend.

“The number of sparse deep learning workloads is rapidly growing, while at the same time, hardware backends are evolving toward accelerating dense operations,” Ye explained. “SparseTIR is flexible by design, enabling it to be applied to any sparse deep learning workload while leveraging new hardware and systems advances.”

In multiple instances, the team found that SparseTIR outperformed highly optimized sparse libraries on NVIDIA HW. Ye and his colleagues earned a Distinguished Artifact Award at ASPLOS ’23, the preeminent conference for interdisciplinary systems research, for their work. 

Based on their scale and the amount of computation required, LLMs are fast becoming one of the most significant hardware workloads — and a potentially significant stumbling block. One of the critical factors for efficient LLM serving is kernel performance on GPUs. To that end, Ye and his collaborators examined LLM-serving operators to identify performance bottlenecks and developed an open-source library, FlashInfer, for enhanced LLM serving using inference acceleration techniques. 

Ye also contributed to Punica, a project led by Ceze and faculty colleague Arvind Krishnamurthy to enable inference of multiple LLMs fine-tuned through low-rank adaptation from a common underlying pretrained model on a single GPU. The team’s approach, which significantly reduces the amount of memory and computation required for such tasks, earned first runner-up in the 2023 Madrona Prize competition.

“Zihao’s work is already having a direct impact,” Ceze noted. “He is a true ML systems star in the making.”

“I’m honored to receive the Graduate Research Fellowship from NVIDIA, which leads the way in research and development of machine learning acceleration. I’m particularly excited to learn from industry experts and to build good systems together for the greater good,” said Ye. 

“I would like to thank Luis, who provided the best guidance and advice, and all my collaborators over the years,” he continued. “UW has a super-collaborative environment where I can team up with people who bring different knowledge and backgrounds, which has greatly expanded my horizons and inspired my research.”

Read more about the 2024 NVIDIA Graduate Research Fellowship recipients on the company’s blog.

Read more →

Melanoma is one of the most commonly diagnosed cancers in the United States. On the bright side, the five-year survival rate for people with this type of skin cancer is nearly 100% with early detection and treatment. And the prognosis could be even brighter with the emergence of medical-image classifiers powered by artificial intelligence, which are already finding their way into dermatology offices and consumer self-screening apps.

Such tools are used to assess whether an image depicts melanoma or some other, benign skin condition. But researchers and dermatologists have been largely in the dark about the factors which determine the models’ predictions. In a recent paper published in the journal Nature Biomedical Engineering, a team of researchers at the University of Washington and Stanford University co-led by Allen School professor Su-In Lee shed new light on the subject. They developed a framework for auditing medical-image classifiers to understand how these models arrive at their predictions based on factors that dermatologists determine are clinically significant — and where they miss the mark.

“AI classifiers are becoming increasingly popular in research and clinical settings, but the opaque nature of these models means we don’t have a good understanding of which image features are influencing their predictions,” explained lead author and Allen School Ph.D. student Alex DeGrave, who works with Lee in the AI for bioMedical Sciences (AIMS) Lab and is pursuing his M.D./Ph.D. as part of the UW Medical Scientist Training Program.

“We combined recent advances in generative AI and human medical expertise to get a clearer picture of the reasoning process behind these models,” he continued, “which will help to prevent AI failures that could influence medical decision-making.”

DeGrave and his colleagues employed an enhanced version of a technique known as Explanation by Progressive Exaggeration. Using generative AI — the same technology behind popular image generators such as DALL-E and Midjourney — they produced thousands of pairs of counterfactual images, which are images that have been altered to induce an AI model to make a different prediction from that associated with the original image. In this case, the counterfactual pairs corresponded with reference images depicting skin lesions associated with melanoma or non-cancerous conditions that may appear similar to melanoma, such as benign moles or wart-like skin growths called seborrheic keratoses. 

The team trained the generator alongside a medical-image classifier to produce counterfactuals that resembled the original image, but with realistic-looking departures in pigmentation, texture and other factors that would prompt the classifier to adjudge one of the pair benign and the other malignant. They then repeated this process for a total of five AI medical-image classifiers, including an early version of an academic classifier called ModelDerm — which was subsequently approved for use in Europe — and two consumer-facing smartphone apps, Scanoma and Smart Skin Cancer Detection.

In order to infer which features contribute to a classifier’s reasoning and how, the researchers turned to human dermatologists. The physicians were asked to review the image pairs and indicate which of the counterfactuals most suggested melanoma, and then to note the attributes that differed between the images in each pair. The team aggregated those insights and developed a conceptual model of each classifier’s reasoning process based on the tendency for an attribute to sway the model towards a prediction of benign or malignant as well as the frequency with which each attribute appeared among the counterfactuals as determined by the human reviewers.

During its audit, the team determined that all five classifiers based their predictions, at least in part, on attributes that dermatologists and the medical literature have deemed medically significant. Such attributes include darker pigmentation, the presence of atypical pigmentation patterns, and a greater number of colors — each of which point to the likelihood a lesion is malignant.

In cases where the classifier failed to correctly predict the presence of melanoma, the results were mixed. In certain instances, such as when the level of pigmentation yielded an erroneous prediction of malignancy when the lesion was actually benign, the failures were deemed to be reasonable; a dermatologist would most likely have erred on the side of caution and biopsied the lesion to confirm. But in other cases, the audit revealed the classifiers were relying not so much on signal but on noise. For example, the pinkness of the skin surrounding the lesion or the presence of hair influenced the decision-making of one or more classifiers; typically, neither attribute would be regarded by dermatologists as medically relevant.

“Pinkness of the skin could be due to an image’s lighting or color balance,” explained DeGrave. “For one of the classifiers we audited, we found darker images and cooler color temperatures influenced the output. These are spurious associations that we would not want influencing a model’s decision-making in a clinical context.”

According to Lee, the team’s use of counterfactual images, combined with human annotators, revealed insights that other explainable AI techniques are likely to overlook.

“Saliency maps tend to be people’s go-to for applying explainable AI to image models because they are quite effective at identifying which regions of an image contributed most to a model’s prediction,” she noted. “For many use cases, this is sufficient. But dermatology is different, because we’re dealing with attributes that may overlap and that manifest through different textures and tones. Saliency maps are not suited to capturing these medically relevant factors.

“Our counterfactual framework can be applied in other specialized domains, such as radiology and ophthalmology, to make AI models’ predictions medically understandable,” Lee continued. “This understanding is essential to ensuring their accuracy and utility in real-world settings, where the stakes for both patients and physicians are high.”

Lee and DeGrave’s co-authors on the paper include Allen School alum and MSTP student Joseph Janizek (Ph.D., ‘22), Stanford University postdoc Zhuo Ran Cai, M.D., and Roxana Daneshjou, M.D., Ph.D., a faculty member in the Department of Biomedical Data Sciences and in Dermatology at Stanford University.

Read the full paper in Nature Biomedical Engineering and a related story by Stanford University. Read more →

When it comes to optimizing the deployment of finite resources in real-world domains, anything less than perfection comes at a price. From agriculture, to e-commerce, to transportation, industries are compelled to optimize everything from processes to personnel — finding the most efficient solution while satisfying a set of constraints that may be at odds with one another.

The optimum solution, then, often requires tradeoffs. But there are some tradeoffs that you just can’t make; it’s physically impossible to assign a quarter of a tractor to a field or half a truck to a route — never mind a third of a flight attendant to an aircraft. Problems such as these, where the only optimal solutions must make use of the whole, call for integer programming.

The integer programming problem, as presented by Karp in his classic 1972 paper, is NP-hard, and the algorithms for solving it run in exponential time. But there are many applications, such as the trio of examples above, that involve a fixed number of variables and thus call for a more limited solution. After all, a farm only has so many acres. For practical reasons, many domains will therefore rely on heuristics that perform well enough but don’t offer the formal guarantees of an exact algorithmic solution. 

But in the decades that followed the release of Karp’s seminal work, researchers have attempted to make progress toward such a solution — and several succeeded. In 1983, mathematician Hendrik Lenstra offered an algorithm with a runtime of 2^{O(n^3)}; four years later, Ravi Kannan and the duo András Frank and Éva Tardos arrived at an n^{O(n)}-time result. After that, progress stalled for more than 30 years.

During that time, as explained in a Simons Foundation article last summer, it remained an open question whether the previous work could be improved upon to exp(O(n)), which would be the fastest achievable under the exponential-time hypothesis. Recently, a duo from the University of Washington — Allen School Ph.D. student Victor Reis and professor Thomas Rothvoss — put such speculation to rest by producing the first (log n)^{O(n)}-time algorithm for solving integer programming problems within a fixed set of variables. 

The pair were motivated to tackle the question after studying the Subspace Flatness Conjecture proposed by Daniel Dadush in 2012, which itself drew upon work by Ravi Kannan and László Lovász in the late 1980’s. Far from falling flat, Reis and Rothvoss’ work earned a Best Paper Award at the 64th IEEE Symposium on Foundations of Computer Science (FOCS 2023) last November in Santa Cruz, California.

“Dadush observed that Kannan and Lovász’ result could be turned into a recursive algorithm for solving integer programs with a run-time of n^n, which yielded a modest improvement over the 1987 result,” explained Rothvoss, who holds a joint appointment in the University of Washington Department of Mathematics. “He further posited that the factor of n could be replaced by a smaller O(\log(n)) term, which would lead to an even more significant improvement. We proved that conjecture to be correct, up to a constant in the exponent.”

To do so, he and Reis employed a combination of well-established and relatively new techniques. One of the tools they were able to draw upon was what Rothvoss termed a “rather deep” result from asymptotic convex geometry known as the ℓ position for symmetric convex bodies, which emerged shortly before the search for a faster integer programming algorithm began in earnest. As described by the duo Tadeusz Figiel and Nicole Tomczak-Jaegermann and by Gilles Pisier, the result stipulates that any symmetric convex set can be linearly transformed to behave, in a certain sense, like a Euclidean ball. This property would prove important to the question at hand when, decades later, Oded Regev and Noah Stephens-Davidowitz produced their Reverse Minkowski Theorem. That theorem, which Rothvoss and Reis also put to good use, implied the correctness of Dadush’s Subspace Flatness Conjecture — assuming the convex set represented in the algorithm is a Euclidean ball.

But whereas the ℓ position allows that high-dimensional slices and projections will have a similar volume as for the ball, there is a huge discrepancy when it comes to low-dimensional slices and projections — precisely that which, according to the work of Kannan and Lovász, one would want to solve for using integer programming. This gave the UW duo a problem to overcome.

“The Euclidean property at the lower dimension was much too weak for Thomas and I to directly apply the work of Regev and Stephens-Davidowitz,” Reis noted. “But we realized we could apply their theorem to the n/2-dimensional sublattice and then recurse on the rest. That gave us a bound for the Subspace Flatness Problem that translated to a significantly improved algorithm for integer programming.”

At present, the result is — you guessed it — theoretical; while it created a buzz in the research community, the duo’s algorithm is unlikely to be put into practice. At least, not yet.

“Sorry, you’re not likely to get your packages any faster based on our paper, but our method could inform the development of future heuristics and open up new questions for us to pursue,” said Rothvoss. “Speaking for myself, I’m perfectly happy with this result — but I know Victor doesn’t like that the exponent isn’t tight.”

“I think it would be really interesting to reduce the original Subspace Flatness Conjecture, with a single logarithmic factor, to other well-studied problems in high-dimensional geometry such as the KLS Conjecture,” Reis said. “There is also the question of whether there could be other approaches for integer programming that don’t rely on subspace flatness. For example, we still don’t know whether the problem can be solved in simply exponential time.”

Read the research paper here and a related Quanta magazine story here. Read more →

In the spring of 2015, the City of Seattle organized a competition and invited teams to “Hack the Commute” by using technology to improve mobility for residents across the city. Among the participants at City Hall was a group of University of Washington students calling themselves Team Hackcessible who had developed a prototype of a web-based trip planning tool called AccessMap. The tool, as envisioned, would combine publicly-sourced and user-submitted data to enable individuals to customize their route around Seattle based on their mobility needs, accounting for factors such as the steepness of hills and the presence or absence of curb cuts.

Under the tutelage of Anat Caspi, director of the Allen School’s Taskar Center for Accessible Technology, Team Hackcessible took home first prize — and the Taskar Center embarked on a journey that would take it from the streets and sidewalks of Seattle, Washington to Santiago, Chile. 

Not long after its triumph at the civic hackathon, the Taskar Center launched OpenSidewalks under the auspices of the UW eScience Institute’s Data Science for Social Good program. OpenSidewalks aims to make data about the public right of way consistent through a publicly available data schema and data collection tools that allow for manual editing in addition to machine learning and automated tools. It also provides educational resources to overcome some of the socio-technical hurdles and misconceptions that people have about access in the public right of way.

With OpenSidewalks up and running, Caspi worked with King County Metro to incorporate the data specification and tooling into the agency’s own paratransit service planning and tools — including data it collects for the provision of paratransit services to people with disabilities in Seattle and across the county. 

“Actual deployment into a large organization revealed a number of important points,” said Caspi. “It became clear to us that people’s understanding of what constitutes ‘accessible’ can vary wildly across transportation planners, administrators and well-meaning crowdsourcing mappers.”

Following the integration of OpenSidewalks into King County Metro’s processes, in 2019 the Taskar Center conducted a study of capacity building. Meanwhile, back on campus, Caspi was pouring her energy into building capacity of a different sort: the capacity to create and use artificial intelligence and data science in a way that addresses, rather than amplifies, ableist bias. For example, in 2017 she launched a vertically integrated project course, Responsible Data Science in Urban Spaces, that offers UW students hands-on experience with developing data-driven software tools to create more accessible, inclusive communities by applying innovation to a range of issues spanning social, economic, health and mobility justice. To date, 241 undergraduate and 23 graduate students have completed the course.

Caspi also leads AI4ALL at UW, a free workshop for high school and first-year college students co-sponsored by the Institute for Foundations of Data Science. Each year, 40 students spend part of the summer learning about how data science, geographic information science and machine learning can be applied to daily life. For current practitioners, Caspi developed a “Non-Ableist AI” workshop and toolkit; since she began offering the workshop four years ago, more than 420 people have participated.

For these and other contributions spanning nearly a decade of leadership at the Taskar Center, Caspi received the 2023 Human Rights Educator Award from the City of Seattle’s Human Rights Commission. 

It just goes to show that, when it comes to advancing accessibility, there’s no place like home. 

“I have always thought making an impact on local communities is where it starts,” said Caspi. “The Seattle commission’s focus on disability rights as civil rights serve as a reminder of the importance of collaborative efforts in creating a more inclusive and just society.”

Those efforts range from the expansive — mapping the accessibility of miles of urban infrastructure — to those of a more human scale. For example, the Taskar Center partnered with not-for-profit Provail Therapy Center to create a library of adapted technology for people with different abilities to borrow for free. The 800 artifacts in the Pacific Northwest Adaptive Technology Library were either donated or created through community education events, where volunteers don protective goggles and get a crash-course in safe use of a soldering iron before adapting battery-operated toys to be switch-accessible for players of different abilities. 

The power of play has been a recurring theme for the Taskar Center. The same year that AccessMap began attracting city leaders’ attention, the center introduced attendees at the Seattle Design Festival to the Universal Play Kiosk. Aligning with the event’s theme of “Design for Equity,” the Universal Play Kiosk demonstrated how to design an immersive environment to engage people of all abilities in collaborative play.

While they have a playful side, Caspi and her colleagues have shown they are serious about the Taskar Center’s motto, “designing for the fullness of human experience.” One of those colleagues — Olivia Quesada, the center’s manager of community engagement and partnerships — accepted the Seattle award on Caspi’s behalf during a ceremony to mark Human Rights Day last month. Quesada completed her UW honors thesis, “Disability Justice for Urban Planners and Designers,” working with Caspi; in her remarks, she shared what makes the center so effective.

“As an interdisciplinary team, we have built a space for recognizing and addressing systemic ableism in various technical and socio-technical systems,” she said. “Our practice of developing, deploying and translating artifacts, both in terms of technology and educational toolkits, is aimed at empowering individuals and communities to confront challenges related to disability justice with a growth mindset.”

The center’s own reach keeps on growing — not only in Seattle, but across the nation and around the globe. In 2021, Caspi launched the Transportation Data Equity Initiative with Co-PI’s Mark Hallenbeck, then director of  the Washington State Transportation Center (TRAC), and iSchool professor Bill Howe, adjunct faculty member in the Allen School and founding co-director of Responsibility in AI Systems & Experiences (RAISE). That initiative builds on the Taskar Center’s previous work to address inequities in transit data and information at scale with support from the U.S. Department of Transportation.

Around the same time, the Taskar Center teamed up with the United Nations advocacy initiative G3ict — short for Global Initiative for Inclusive Information and Communication Technologies — on AI for Inclusive Urban Sidewalks. The project combines artificial intelligence with on-the-ground community partnerships to improve pedestrian accessibility in cities around the globe, with support from Microsoft’s AI for Accessibility grant and the Open Data Campaign. In 2022, the Taskar Center and G3ict earned the SmartCity Expo World Congress Living & Inclusion Award, one of a set of honors recognizing pioneering initiatives and ideas to make cities around the world more livable, sustainable and economically viable. 

Which brings us back to where it all began. In the years following Hack the Commute, members of the original AccessMap team — including Allen School postdocs Nick Bolton, Ricky Zhang, and Sachin Mehta, all graduates of the UW Department of Electrical & Computer Engineering — and a succession of new student researchers drove the project forward. In 2017, Caspi and the team introduced their AccessMap web-based tool for the public that, upon first release, offered personalized trip planning for the Washington cities of Seattle, Mt. Vernon and Bellingham. Fast forward almost seven years later, and the center released a new, expanded version called AccessMap Multimodal. The latest iteration incorporates indoor transit information, where available, along with sidewalk data and extends to 11 cities worldwide — including several participants in the aforementioned AI for Inclusive Urban Sidewalks project. After racking up 65,000 user routing requests, AccessMap and its user base continues to grow. 

Whether in Santiago or Seattle, Caspi’s tireless efforts at advocacy and education have put the Taskar Center itself on the map. But she’s eager to share the plaudits with her collaborators.

“This recognition is not just a testament to my efforts but a celebration of the collective dedication to promoting disability human rights and inclusivity by the Taskar Center,” Caspi said.

For more inspiration, join the Taskar Center today (Thursday, January 25) for the Ben Taskar Annual Memorial Events (12:00–3:00 pm) — including an overview of AccessMap Multimodal — and the Ben Taskar Memorial Distinguished Lecture (3:30–5:00 pm) featuring Drago Anguelov, vice president and head of research at Waymo.

Read more →

Allen School professor and alum Leilani Battle (B.S., ‘11) originally wanted to be a game developer. As a kid growing up in Bremerton Washington, Battle saw a glimpse of her future every time she booted up her family’s Nintendo 64. Whether dodging shells and banana peels in Mario Kart or catching them all as a Pokemon trainer, she saw how imagination could manifest itself in new and inventive ways. 

“I loved immersing myself in the worlds created by others through video games,” she said. “I saw games as a nice mix of creativity and problem solving. Computer science seemed like a sensible step towards this childhood dream.”

At the University of Washington, Battle’s interests shifted. The creative problem-solving spark remained, but instead of immersing herself in games she immersed herself in data — specifically, new and improved ways to explore the vast quantities available to scientists and analysts. Battle went on to earn her Ph.D. from MIT before returning to the Allen School to complete a postdoc in the Interactive Data Lab and UW Database Group. After spending three years as a professor at the University of Maryland, College Park, Battle returned once again to the Allen School in 2021 — this time as faculty — to co-lead the Interactive Data Lab with colleague Jeffrey Heer. 

Drawing upon techniques from databases, human-computer interaction and visualization, Battle is not playing around when it comes to her current research focused on modeling user behavior to not only understand but optimize users’ ability to glean actionable insights from their data. She has earned a string of professional accolades for this work; the most recent came in October, when she received the 2023 VGTC Visualization Significant New Researcher Award from the IEEE Visualization and Graphics Technical Community for her contributions to “interactive data-intensive systems for exploratory data analysis.”

As one example of her contributions, last year Battle and collaborators at the University of Maryland examined how users incorporate languages like D3 into their implementation workflows when designing visualizations. They performed a mixed methods analysis of nearly 38,000 posts on Stack Overflow — a popular resource for D3 users — and noted that a gap exists between how creators of data visualizations conceptualize their designs versus how they reason about D3’s code structure. The authors found that the resulting disruption to their workflows discouraged more widespread adoption of these tools. Battle and her colleagues proposed multiple approaches for ameliorating these issues, including smoother integration of languages like D3 with other visualization tools, automating the process for generating example design galleries, and improving D3’s support infrastructure to enable users to more easily find and incorporate meaningful code components into their workflows.

“When visualization languages are developed and tested in a vacuum, without considering how and where people use them, they can fall short of addressing those users’ needs,” said Battle, who alongside her co-authors earned the award for Best Short Paper at last year’s IEEE Visualization and Visual Analytics Conference (VIS 2022). “By being mindful of how users interact with these tools in practice as they are developed, we can improve users’ information access and empower them to explore a broader range of effective visualization designs.”

Battle and another group of University of Maryland colleagues subsequently took a similar tack in an effort to understand how users approach sensemaking, an iterative process through which they refine their visualizations to deepen understanding of their data. This time, they applied their mixed methods analysis to more than 2,500 Jupyter notebooks — a popular tool for documenting the sensemaking making process in data science — on Github to understand how the sensemaking pipeline evolves over time in order to better support a variety of sensemaking activities, such as annotation, branching analysis and documentation. The team earned a Best Paper Honorable Mention at the Association for Computing Machinery’s Conference on Human Factors in Computing Systems (CHI 2023) this past spring for their work.

More recently, Battle took a deep dive into historical whaling data as part of Computing for the Environment, a cross-campus initiative aimed at applying interdisciplinary research in computing, engineering and environmental sciences to address challenges ranging from climate change to wildlife conservation. She and Allen School Ph.D. student Ameya Patil teamed up with Trevor Branch, a professor in the UW School of Aquatic & Fishery Sciences, and Zoe Rand, a Ph.D. student in the UW’s Quantitative Ecology and Resource Management (QERM) program, to develop WhaleVis, an interactive dashboard that enables scientists to explore roughly a century’s worth of historical data maintained by the International Whaling Commission to inform current whale conservation efforts — without consuming an onerous amount of computing resources. 

“Scientific data is a really important aspect of big data, but scientists all over the world have access to completely different hardware and software. Maybe they can’t use big servers to process huge data sets quickly,” Battle explained to UW News. “So when creating WhaleVis we had to ask: How do we design a tool that can visualize millions of data points, but that doesn’t rely on super beefy servers?”

Battle and her colleagues presented WhaleVis at IEEE VIS 2023 held in Melbourne, Australia this fall — the very same conference at which she collected the VGTC recognition. The award follows her selection as a Sloan Research Fellow earlier this year, after having earned the TCDE Rising Star Award and a National Science Foundation CAREER Award in 2022. Battle previously was named among MIT Technology Review’s Innovators Under 35 for her earlier work on projects such as ForeCache, a system for reducing latency in large-scale data exploration by prefetching data based on user behavior.

Read the full VGTC award citation here.

Roger Van Scyoc contributed to this story.

Read more →

Misinformation can spread like wildfire on social media, fueled in part by platforms’ tendency to prioritize engagement over accuracy. This puts the onus on individual users to determine the veracity of posts they see and share on their feed. Likewise, when it comes to violence, profanity and other potentially harmful content, users are often left to fend for themselves in the face of indifferent or inadequate moderation. The current state can make social media platforms a harrowing place — particularly for members of marginalized communities.

Researchers in the University of Washington’s Social Futures Lab led by Allen School professor Amy X. Zhang hope to change that by designing social media tools that empower users while minimizing the burden of managing their online experiences.

“A big problem to me is the centralization of power — that the platforms can decide what content should be shown and what should get posted to the top of one feed for millions of people. That brings up issues of accountability and of localization to specific communities or cultures,” Zhang explained in an interview with UW News. “I’ve been looking at what it would mean to decentralize these major platforms’ power by building tools for users or communities who don’t have lots of time and resources.”

At the 26th ACM Conference on Computer-Supported Cooperative Work And Social Computing (CSCW 2023) last month, Zhang and her co-authors shared promising results from two such projects: an exploration of provenance standards to help gauge the credibility of media, and the creation of personal moderation tools to help stem the tide of harmful content.

Seeing is believing?

The proliferation of misinformation on social media platforms has led to a “credibility crisis” when it comes to online content — including that posted by local and national news organizations. Visual content, in particular, can be used to manipulate users’ understanding of and reaction to events. While reverse-image search allows users to investigate potentially problematic content, this approach has its limitations; the results are often noisy or incomplete, or both.

Lab member Kevin Feng, a Ph.D. student in the UW Department of Human Centered Design & Engineering, believes the introduction of provenance standards could provide a pathway to restoring trust in what he calls the “information distribution infrastructure.”

“Misinformation can spread through social networks much faster than authentic information. This is a problem at a societal scale, not one that’s limited to a particular industry or discipline,” said Feng, lead author on the CSCW paper. “Being able to access a piece of media’s provenance, such as its prior edit history and sources, with just a click of a button could enable viewers to make more informed credibility judgments.”

Feng turned to the Coalition for Content Provenance and Authenticity (C2PA), which at the time  was in the midst of developing an open-source technical standard for media authoring tools such as Adobe Photoshop to embed a distinct signature into an image’s metadata every time someone edits it. The goal was to provide a verifiable chain of provenance information — essentially, a detailed edit history — to viewers of online content. 

But given the overall erosion of trust in online media, an important question remained.

“Even if we can reliably surface provenance information, we still didn’t know if or how that would impact users’ credibility judgements,” Feng noted. “This is an important question to answer before deploying such standards at scale.” 

Seeking answers, Feng and Zhang teamed up with Nick Ritchie, user experience design principal at the BBC, and Pia Blumenthal and Andy Parsons, lead product designer and senior director, respectively, of Adobe’s Content Authenticity Initiative. The team ran a study involving 595 participants in the United States and United Kingdom in which they measured how access to provenance information altered users’ perceptions of accuracy and trust in visual content shared on social media.

The team developed dual Twitter-esque social media feeds for the study: one regular feed, and one containing provenance information accessible through user interfaces built in accordance with the C2PA standard. The team relied on pre-existing images and videos sourced mainly from the Snopes “fauxtography” archives, representing a mix of truthful and deceptive content. Participants were asked to view the same series of images on the control feed followed by the experimental feed and rate the accuracy and trustworthiness of each piece of content on a 5-point scale. This enabled the researchers to gauge how each user’s perception of media credibility shifted once they had access to information about the content’s provenance.

And shift, it did.

“With access to provenance information, participants’ trust in deceptive media decreased and their trust in truthful media increased,” Feng reported. “Their ability to evaluate whether a claim associated with a particular piece of media was true or false also increased.”

Feng and his co-authors found that the use of provenance information comes with a couple of caveats. For example, if an image editor is incompatible with the C2PA standard, any changes to the image may render the chain of provenance information “incomplete,” which does not speak to the nature of the edit. In addition, if a malicious actor attempts to tamper with the metadata in which the provenance information is stored, the provenance is rendered as “invalid” to warn the user that suspicious activity may have occurred — whether or not the attempt was successful. 

The team discovered that such disclosures had the effect of lowering trust in truthful as well as deceptive media, and caused participants to regard both as less accurate. The researchers were also surprised to learn that many users interpreted provenance information as prescribing a piece of media’s credibility — or lack thereof.

“Our goal with provenance is not necessarily to prescribe such judgements, but to provide a rich set of information that empowers users to make more informed judgments for themselves,” Zhang notes. “This is an important distinction to keep in mind when developing and presenting these tools to users.”

The team’s findings informed the eventual design of the production version of the C2PA standard, Content Credentials. But that’s not the only distinction that is reflected in the standard. When the study launched, generative AI had not yet come into the mainstream; now, AI images seem to be everywhere. This poses important questions about disclosure and attribution in content creation that, in Feng’s view, make provenance standards even more timely and relevant.

“As AI-generated content inevitably starts flooding the web, I think that provable transparency — being able to concretely verify media origins and history — will be crucial for deciphering fact from fiction,” he said.

Everything in moderation

When problematic posts cross the line from mendacious to mean, there are a variety of approaches for reducing users’ exposure to harmful content. These vary from platform–wide moderation systems to user-configured tools based on personal preference in response to the content posted by others. The emergence of the latter has sparked debate over the benefits and drawbacks of putting moderation decisions on individual users — attracting fans and foes alike.

In their recent CSCW paper, lead author Shagun Jhaver and his colleagues investigated this emerging paradigm, which they dubbed “personal content moderation,” from the perspective of the user.

“With more and more people calling for greater control over what they do, and do not, want to see on social media, personal moderation tools seem to be having a moment,” explained Jhaver, a former Allen School postdoc who is now a professor at Rutgers University where he leads the Social Computing Lab. “Given the growing popularity and utility of these tools, we wanted to explore how the design, moderation choices, and labor involved affected people’s attitudes and experiences.”

Personal moderation tools fall into one of two categories: account-based or content-based. The former includes tools like blocklists, which prevent posts by selected accounts from being displayed in a user’s feed. Jhaver, Zhang and their co-authors — Allen School postdoc Quan Ze Chen, Ph.D. student Ruotong Wang, and research intern Alice Qian Zhang, a student at the University of Minnesota who worked with the lab as part of the Computing Research Association’s Distributed Research Experiences for Undergraduates (DREU) program — were particularly interested in the latter. This category encompasses a range of tools enabling users to configure what appears in their feed based on the nature of the content itself.

The researchers built a web application that simulates a social media feed, complete with sample content and a set of interactive controls that can be used to reconfigure what appears in the feed. The tools included a word filter, a binary toxicity toggle, an intensity slider ranging from “mildly” to “very” toxic, and a proportion slider for adjusting the ratio of benign to toxic comments. They then enlisted a diverse group of two dozen volunteers to interact with the configuration tools and share their insights via structured interviews.

The team discovered that participants felt the need to build a mental model of the moderation controls before they could comfortably engage with the settings, particularly in the absence of robust text explanations of what various moderation terms meant. Some users switched back and forth between their settings and news feed pages, using examples to tweak the configuration until they could verify it achieved their desired goals. The test interface supported this approach by enabling participants to configure only one of the four moderation interfaces at a time, which meant they could observe how their choices changed the feed.

This seemed particularly helpful when it came to slider-based controls that categorize content according to high-level definitions such as “hateful speech” or “sensitive content” — categories that users found too ambiguous. Context is also key. For example, participants noted that whether profanity or name-calling is offensive depends on the intent behind it; the mere presence of specific words doesn’t necessarily indicate harm. In addition, some participants were concerned that filtering too aggressively could curtail the visibility of content by members of minority communities reclaiming slurs as part of in-group conversations.

“Whether at the platform level or personal level, users crave more transparency around the criteria for moderation. And if that criteria is too rigid or doesn’t allow for context, even user-driven tools can cause frustration,” Zhang said. “There is also a tradeoff between greater agency over what appears in one’s feed and the effort expended on configuring the moderation settings. Our work surfaced some potential design directions, like the option to quickly configure predetermined sets of keywords or enable peer groups to co-create shared preferences — that could help strike a balance.”

Another tension Zhang and her team observed was between content moderation policies and users’ desire to preserve freedom of speech. Although users still believe there is a role for platforms in removing the most egregious posts, in other instances, personal moderation could provide an attractive alternative to a top-down, one-size-fits-all approach. In what Jhaver terms “freedom of configuration,” users choose the nature of the content they want to consume without curtailing other users’ self-expression or choices.

“These tools are controlled by the user, and their adoption affects only the content seen by that user,” Jhaver noted. “Our study participants drew a distinction between hiding a post from their individual feed and a platform removing a post for everyone, which could be considered censorship.”

While it may not chill free speech, personal moderation could meet with an icy reception from some users due to another social media phenomenon: the dreaded FOMO, or “fear of missing out.”

“Many of our participants worried more about missing important content than encountering toxic posts,” explained Jhaver. “Some participants were also hesitant to suppress inappropriate posts due to their desire to remain informed and take appropriate action in response.”

Whether it has to do with moderation or misinformation, the researchers are acutely aware of the societal implications of their work.

“As social computing researchers, we are the architects of our digital spaces,” Feng said. “The design decisions we make shape the ways with which people interact online, who they interact with, how they feel when doing so, and much more.”

For more on this topic, see the UW News Q&A featuring Zhang here.

Roger Van Scyoc contributed to this story. Read more →

New approaches to finetuning large language models that decrease computational burden while enhancing performance. A robotic arm that safely delivers a forkful of food to someone’s mouth. A system that combines wireless earbuds and algorithms into a low-cost hearing screening tool.

These are just a sample of the nearly 60 projects that were on display during the Allen School’s Research Showcase and Open House at the University of Washington last week, capping off a day-long celebration of computing innovations that are advancing the field and addressing societal challenges. Nearly 300 Industry Affiliate partners, alumni and friends participated in the 2023 event, which included sessions devoted to computer science ethics, intelligent transportation, computing for sustainability, computing for health, natural language processing and more.

Everybody’s talking about LLMs

Attendees got the chance to sink their teeth into some of the latest advances in natural language processing during a luncheon keynote by Allen School professor Hannaneh Hajishirzi exploring the science behind large language models and the development of models to serve science.

“We have witnessed great progress in large language models over the past few years. These models create extremely fluent text — conversation-like text — and also code,” said Hajishirzi, who holds the Torode Family Professorship at the Allen School and is also senior director of AllenNLP at the Allen Institute for AI. ”Now they are being deployed in a diverse range of applications. And everybody these days is talking about their impact on society, their risks, their economic impacts, and so on.”

Those impacts and risks leave plenty of open questions for AI researchers to resolve, as LLMs continue to be computationally expensive, error-prone and difficult to maintain. They are also largely being developed by private companies.

“All of these models are proprietary,” Hajishirzi noted. “So it’s very hard for AI researchers to actually understand and analyze what is going on.”

Hajishirzi and her colleagues favor a more open approach to building models that are transparent, reproducible and accessible. But there are many definitions of “open.” Even if the company opens up the API or makes a model available for research purposes, restrictions remain — such as the inability to access the data on which the models are trained.

As an alternative, Hajishirzi and her collaborators created OLMo, short for Open Language Model. OLMo is a full language modeling pipeline in which “everything is open,” from pre-training to reinforcement learning through human feedback (and all stages in between). By being so transparent and engaging the broader AI research community, Hajishirzi hopes the project will help narrow the gap between the public and private sectors. Their good intentions are not limited to advancing AI research, either; the team is also developing the capability to advance scientific discovery in other disciplines by fine tuning and training on their data.

To that end, Hajishirzi and her colleagues developed a large-scale, high-quality pretraining dataset cleverly named Dolma, short for “data to feed OLMo’s appetite.” The dataset, which comprises 3.1 trillion tokens in total, is significantly larger than previous open datasets. A significant portion — 2.6 trillion tokens — is web data covering diverse domains, from Reddit to scientific data, filtered to eliminate toxicity and personally identifying information as well as duplication. Dolma has been downloaded 320,000 times in just the past month.

But how does this approach compare to that of state-of-the-art closed models? When it comes to the latter, “there are too many question marks,” Hajishirzi noted, pointing out that we don’t have sufficient information about the datasets — including not knowing how many tokens the models are trained on. 

That is not a problem when it comes to the work of Hajishirzi and her collaborators — including the development of novel approaches to instruction tuning to enable pretrained models to generalize to new applications. Hajishirzi described the result of those efforts, a project called Tülu, as “the largest, best and open instruction tuned model at this point.” And the team continues to make improvements; for example, they have added the ability to extract information from scientific papers and to perform parameter-efficient finetuning for use in low-resource contexts. The researchers have also developed an effective evaluation framework that includes in-loop evaluation of the training at every step of the process. 

Such progress does not come without a cost, however.

“This project required a lot of compute. It still requires a lot of GPUs and compute,” Hajishirzi observed, citing the need to improve computational efficiency so that more communities can make use of these models.

How low can you go?

As it happens, multiple Allen School researchers are attempting to answer this question — and answer Hajishirzi’s call — by exploring techniques for making LLMs more efficient. Teams shared their results from projects addressing this and a range of other challenges during the open house and poster session.

The event culminated with Scott Jacobson, managing director at Madrona Venture Group, announcing the recipients of the Madrona Prize, which highlights cutting-edge research at the Allen School with commercial potential. In his remarks, Jacobson highlighted the firm’s long standing partnership with the Allen School, which extends to supporting multiple startup companies based on student and faculty research that is helping to shape the future of the field.

”There’s so much great research here” said Jacobson. “Over the years, a number of themes that I think are now kind of commonplace in tech were really pioneered here. A lot of those themes you’ve seen in the poster session on Industry Affiliates day — cloud computing, edge computing, computer vision, a lot of applied machine learning and AI. And so it’s just really fun every year for us to get the opportunity to do this.”

Madrona Prize winner/ QLoRA: Efficient Finetuning of Quantized LLMs

Allen School Ph.D. student Tim Dettmers accepted the grand prize for QLoRA, a novel approach to finetuning pretrained models that significantly reduces the amount of GPU memory required — from over 780GB to less than 48GB — to finetune a 65B parameter model. With QLoRA, the largest publicly available models can be finetuned on a single professional GPU, and 33B models on a single consumer GPU, with no degradation in performance compared to a full finetuning baseline. The approach will help close the gap between large companies and smaller research teams, and could potentially enable finetuning on smartphones and in other low-resource contexts. The team behind QLoRA includes Allen School Ph.D. student Artidoro Pagnoni; alum Ari Holtzman (Ph.D., ‘23), incoming professor at the University of Chicago; and professor Luke Zettlemoyer, who is also a research manager at Meta.

Madrona Prize First Runner Up / Punica: Multi-Tenant LoRA Fine-tuned LLM Serving

Another team earned accolades for their work on Punica, a framework that makes low-rank adaptation of pre-trained models for domain-specific tasks more efficient by serving multiple LoRA models in a shared GPU cluster. Punica’s new CUDA kernel design allows for batching of GPU operations for different models while requiring a GPU to hold only a single copy of the underlying pre-trained model — significantly reducing the level of memory and computation required. The research team includes Allen School Ph.D. students Lequn Chen and Zihao Ye; Duke University Ph.D. student Yongji Wu; Allen School alum Danyang Zhuo (Ph.D., ‘19), now a professor at Duke; and Allen School professors Luis Ceze and Arvind Krishnamurthy.

Madrona Prize Second Runner Up / Wireless Earbuds for Low-cost Hearing Screening

Allen School researchers were recognized for their work with clinicians on OAEbuds, which combines low-cost wireless acoustic hardware and sensing algorithms to reliably detect otoacoustic emissions generated by the ear’s cochlea. The system offers an alternative to conventional — and expensive — hardware to make hearing screening more accessible in low- and middle-income countries. Allen School Ph.D. student Antonio Glenn accepted on behalf of the team, which also includes Allen School alum Justin Chan (Ph.D., ‘22), incoming professor at Carnegie Mellon University; professors Shyam Gollakota and Shwetak Patel, who has a joint appointment in the UW Department of Electrical & Computer Engineering; ECE Ph.D. student Malek Itani; Drs. Randall Bly and Emily Gallagher of UW Medicine and Seattle Children’s; and audiologist Lisa Mancl, affiliate instructor in the UW Department of Speech & Hearing Sciences.

People’s Choice Award / ADA, the Assistive Dexterous Arm: A Deployment-Ready Robot-Assisted Feeding System

Also affectionately referred to as “the food thing” to attendees who overwhelmingly voted it their favorite demo of the night, ADA aims to address a variety of technical and human challenges associated with robot-assisted feeding to improve quality of life for people with mobility limitations. The researchers invited visitors to try the system for themselves by using a smartphone app to direct ADA in feeding them forkfuls of fruit. Ph.D. student Amal Nanavati accepted the award from professor Shwetak Patel, the Allen School’s associate director for development and entrepreneurship. The team also includes Ph.D. students Ethan Gordon and Bernie Hao Zhu; undergraduate researcher Atharva Kashyap; Haya Bolotski, a participant in the Personal Robotics Lab’s youth research program; Allen School alum Raida Karim (B.S., ‘22); postdoc Taylor Kessler Faulkner; and professor Siddhartha Srinivasa. Read more about the robot-assisted feeding project in a recent UW News Q&A with the ADA team here.

For more about the Allen School’s 2023 Research Showcase and Open House, read GeekWire’s coverage here and the Madrona Prize announcement here.

Read more →

It can feel lonely being the first in your family to pursue a four-year degree.

How do you apply? How will you pay for it? What major should you choose? How will you navigate your new surroundings, not to mention make new friends? If you run into difficulty, where do you turn for help?

And what are “office hours,” anyway?

Nearly one-third of the more than 43,000 undergraduates enrolled at the University of Washington are first-generation. So while it may seem lonely at times, they are not alone. To remind them of this fact — and to remind everyone at UW of the many ways in which first-gen students enrich our campus community — each year on November 8th the University participates in the National First-Generation College Celebration. To highlight the Allen School’s diverse first-gen community, we asked students to share what it means to them to “be the first” and any wisdom they have for those who have yet to embark on their first-gen journey.

Zander Brumbaugh: Turning a hobby into an opportunity for connection and empowerment

While Zander Brumbaugh knew from a young age that he wanted to be a scientist, he didn’t know he wanted to be a computer scientist. Growing up in Tumwater, Washington, Brumbaugh turned his fascination with the inner workings of various systems into a hobby making video games in high school and, eventually, the beginnings of a career in computing. Currently pursuing his master’s degree in the Allen School’s combined B.S./M.S. program, Brumbaugh has refined his career goals to focus on artificial intelligence research — an “immensely important field” in which he hopes to make a positive impact. To start, he is writing a book on how to adapt and use language models effectively for specific needs as a way of promoting public literacy around these rapidly emerging technologies.

What does it mean to you and/or your family to be the first to pursue a bachelor’s degree?

My family has always been very supportive of the work that I do and my decision to pursue higher education. Both of my parents were unable to attend college due to financial limitations. Because of the scholarships I received, I was able to overcome this and be the first in my family to earn a degree, for which I am eternally grateful. In short, my degree is a source of empowerment; it gives me the ability to create new opportunities for myself and to connect with like-minded individuals who share similar goals.

What has been the most challenging aspect of being a first-gen student?

College is quite different from high school or anything else most students are likely to have encountered in their early academic careers. Finding the groove in my first few quarters wasn’t easy, especially with the start of the COVID-19 pandemic less than halfway into my first year. I made a group of close friends who came from many different backgrounds, and I found my experience improved greatly. Being a first-gen student, I didn’t have anyone at first to help me navigate student life, but we ultimately found our way through it together. 

And the most rewarding?

By far the most rewarding part is simply being able to call home and tell my parents what I’ve been up to. Both of my parents are retired and enjoy hearing the details of my classes, activities with friends, and my research — though they say most of it goes over their heads! My father is my biggest promoter; oftentimes I’ll receive messages online from people who work as cashiers at stores or waitstaff at restaurants whom he’s cheerfully told about my books and games. My parents’ pride inspires me to be the best version of myself — and also thankful for the opportunities I’ve been given, knowing it’s something they didn’t have.

What motivated you to continue on and get your master’s?

As it was always my goal to become a researcher, I began looking for undergraduate research opportunities during my sophomore year. I first worked in AI for vision and language for creative applications and eventually found intersections with robotics that greatly interested me. I joined the ARK lab led by professor Noah Smith during my senior year and started my journey with natural language processing (NLP) research. Wanting to continue my research and eventually pursue a Ph.D. in the future, I applied to the B.S./M.S. program and was accepted. So far, the experience has been everything I imagined; the program provides an environment where I’m immersed in intriguing research, exchanging ideas with others both in and outside of my field and developing projects across various topics.

What advice would you give to other first-gen students?

Heading off for college is an exciting time, full of new experiences. Even if you have family or friends who went to college, it can be difficult to find advice on how exactly you should be approaching different problems — and if you don’t, it can be even more so. While everyone’s experience may be different, finding even a small, close group of friends can help to make a support system. You can help each other navigate your classes, work, or simply your social lives. I would encourage you to check out a club meeting, be outgoing whenever you can, and try to look for others with whom you might share something in common (or not!). There are also mentorship programs offered by multiple groups affiliated with the Allen School that may be helpful in getting you started.

Daniel Campos Zamora: Making a career out of making change and helping people at scale

Daniel Campos Zamora followed what he calls a “long and winding road” to computer science that extends back to his birthplace of Costa Rica. Growing up in New Jersey, Campos Zamora had always been interested in making interesting things; he just never considered making a career out of it. That changed after he began an interdisciplinary degree in psychology and art at Carnegie Mellon University. There, he discovered programming tools like Arduino and Processing that made him realize how powerful computing could be as a medium for change. After earning his bachelor’s, Campos Zamora worked for a professor of human-computer interaction, and later, for Disney Research; that combination of experiences caused him to realize that he wanted to do HCI research himself. The road eventually led him to pursue a Ph.D. in the Allen School’s Makeability Lab working professor Jon Froehlich — and to tap into his first-gen experience in his roles as a reviewer for the school’s Pre-Application Mentorship Service (PAMS) and faculty recruiting.

What does it mean to you and/or your family to be the first to pursue a bachelor’s degree?

I was raised by a single mom who immigrated to the U.S. because she had high hopes for us to get an education and get ahead. She always instilled in me and my siblings that she wanted us to go to college, but she didn’t really understand what that entailed. College didn’t really feel real to us; no one in our family had gone to college, and we didn’t know a lot of people in this country who had college degrees. My mom maybe took one class back in Costa Rica before she had to drop out to have my brother. So it meant the world to her that her three kids were able to get degrees. After graduating, I framed my diploma and gave it to her as a Christmas gift, because I knew how much it meant to her. And if you don’t know, the CMU diploma is gigantic!

What was the most challenging aspect of being a first-gen student?

I think the whole experience of college is really different when you’re first generation. I was the only one of us to move away for college and be away from family and live in the dorms. When you do that, you don’t have a support system, and you don’t know what kind of support is available at school. You don’t know about office hours; you may know they exist, but you don’t know what they actually mean, and you don’t know what any of the offices on campus mean or do. You don’t know what you don’t know yet — you’re dealing with “unknown unknowns.” I struggled because I didn’t identify with a lot of people at my institution, and it was really hard finding support.

How did you navigate those “unknown unknowns”?

I went back to what I knew: I can just work on the classes and do my best. I think I stumbled through some of the other parts, like eating, taking care of yourself, your mental health. You don’t realize, when you’re away from your family, it requires extra work to do that. The saving grace for me was that I met someone on my floor who has a similar background but who knew the school better after doing a summer program there. Through him I got involved in a minority organization and found a support system. And they led me to the campus resource center that assists minoritized students, including first-gen and low income. Through their counselors, I got a lot of support — but it took two years before I even knew that office was there.

What was the most rewarding aspect of being a first-gen student?

I feel like I’m in a much better place to help younger family members thinking about college to understand what it actually means to go down this path. I can take them to look at colleges and help them understand the processes and that there is so much more to the college experience than just getting good grades.

For me personally, what was most rewarding was being exposed to really talented people and really exciting ideas, and to be able to take advantage of resources once I knew they were available. I think it opened up a lot of doors for me, and I would not be at the Allen School if it was not for that experience at CMU. Also, the friends and connections that I made there — I know I have those relationships for a lifetime.

What advice would you give to aspiring first-gen students?

Advocate for yourself, but also be able to admit that you don’t know stuff. I feel that when you get to college, the imposter syndrome — that feeling like, “I don’t belong here” — is so aggravated because you are the first one. So I think it’s knowing that you do belong, but you might need help. Also, usually people who make it to these schools have done well academically, and they might not have struggled too much up until that point. And because so much importance is placed on going to college and getting good grades, when you do run into roadblocks, it’s so disorienting and discouraging.

It takes a lot of courage to acknowledge that you’re struggling and to ask for help. I think that’s very tough for first-gen students who may not even know that they are struggling. I wish someone had told me that it’s okay to ask for help. I’ve had to ask for help here doing my Ph.D. — I’m the first in my family to go to grad school, so that’s a totally new thing. The more you ask for help, and the earlier you do that, the better off you’ll be.

Ha Vi Duong: Choosing her own path while embracing the power of creative problem solving

As a high school student in Moses Lake, Washington, Ha Vi Duong envisioned a career in medicine. While she soon realized that she wanted to do something else with the rest of her life, she wasn’t sure what that something was. A conversation with an advisor — and an encounter with programming through Girls Who Code — helped her see how computer science would allow her to exercise her creativity while solving real-world problems. Duong entered the Allen School as part of the 2021-22 cohort of Allen Scholars. She later took on the role of chair of GEN1, a student group dedicated to empowering and guiding first-gen students, and joined the Vietnamese Student Association’s VSAUW Dance Team. Throughout her time at UW, she has been determined to work hard not just for her own future, but also for that of her parents — in appreciation for the sacrifices they’ve made.

What does it mean to you and your family to be among the first to pursue a bachelor’s degree?

My parents always emphasized the importance of education and believed it should be a top priority in life. I remember them sharing stories about their own experiences when they had to help provide for their families instead of continuing their education. For them, education didn’t always come first. They made the selfless decision to come to America in search of a better life, especially for their children. They’ve worked tirelessly to make this happen, running a restaurant that demands long hours and hard work. 

One summer, I got a firsthand look at what they go through when I helped out at their restaurant. It was an eye-opener, and I gained a deep appreciation for what my parents do every day. When I talked to them about how tough it is, they told me something that has stuck with me ever since: “You should work hard to not have a job like ours. We didn’t know what else to do.” It made me realize how lucky I am. I have the chance to pursue higher education, to explore many career options, and choose my own path. This awareness has made me incredibly grateful.

What has been the most challenging aspect of being a first-gen student at the Allen School?

The most challenging aspect has been dealing with imposter syndrome. The rigorous coursework often makes it feel as though I’m behind compared to my peers. To this day, I still can’t believe that I am able to be where I am. However, amidst this struggle, I’ve discovered a valuable support system within the Allen School community and an understanding that we are all on our own paths and are here for a reason, which has been key in overcoming these challenges. 

What about the most rewarding?

The most rewarding part of my experience has been the sense of accomplishment that comes from successfully completing those demanding courses. It’s immensely satisfying to see the progress I’ve made and to be able to connect the material learned in one class to another and eventually apply this knowledge in the real world. This interconnectedness between academic learning and practical application makes the educational journey at the Allen School both challenging and deeply fulfilling.

Any advice for other first-gen students at UW?

For first-gen students, it’s essential to remember that UW offers fantastic programs and a wide range of groups and organizations. While it may initially feel overwhelming, it’s all about the effort you put into discovering resources and building connections to support you in your college journey. Don’t hesitate to step out of your comfort zone and make connections; you never know where it might lead! It’s normal to feel a bit lost, but remember that everyone has their unique path and pace.

What are you hoping to do after graduation?

I’m currently in the process of exploring my post-graduation options, and I believe that finding a career where I can witness the tangible impact of my work is crucial. While I don’t have a specific plan in place just yet, I’m actively seeking opportunities that align with my interests and values. My goal is to pursue a path that allows me to make a meaningful difference in the world and see the results of my efforts come to life.

Derrik Petrin: Rediscovering his love of computing by leaving the lab and entering the arena

Derrik Petrin went all the way to Yale University to earn a bachelor’s degree in biochemistry before he realized that he did not, in fact, enjoy working in a lab. As a middle-school student in Issaquah, Washington, he had taken a programming class and liked it; he also liked the original Magic the Gathering card game by Wizards of the Coast. After he returned to this coast, Petrin eventually parlayed both into a position with the company as a software development engineer after spending some time as a freelance software consultant. Lately, Petrin has been working on the team that produces the digitized version of the game, Magic the Gathering Arena. When he’s not getting paid to play during work hours, he’s advancing his own story arc by pursuing a graduate degree in computer science through the Allen School’s flexible Professional Master’s Program (PMP) — which offered Petrin not only the opportunity to obtain a computer science degree but also to explore what his next chapter might be. He also has connected with his roots through his involvement in the local Hungarian-American Association.

What did it mean to you and/or your family to be among the first to pursue a bachelor’s degree?

It’s funny — I didn’t really start thinking of myself as first-generation until near the end of undergrad. I went to school on the Eastside, on the plateau, and didn’t really appreciate the differences. Some of my aunts and uncles went to college, but my dad’s parents didn’t, my dad didn’t, my mom emigrated from Hungary. So it didn’t really start dawning on me until I was in undergrad, when I realized that all of my classmates’ parents were professionals — lawyers, doctors, engineers and mathematicians — and mine weren’t. I started noticing how my parents didn’t have any advice they could give me. I always thought I would get a degree; I think my parents always expected that, too. So it was hard to imagine me not doing that. 

What was the most challenging aspect of being a first-gen student?

My parents are not academically inclined. By late middle school and high school, I felt really comfortable in an academic setting and was used to planning and deciding everything myself. One of the reasons I ended up going to Yale was that it had the most generous financial aid package. I remember in my senior year seeing a statistic about the percentage of students who received no financial aid at all — and that’s a pretty high income cut-off — and it was a large number of students. And then I noticed that a lot of the classmates that I had formed close friendships with were also on some form of financial aid. So it dawned on me that this stuff tends to organically group us together without us realizing it.

I also had some pretty severe mental health struggles. Yale has had some publicity in recent years about their poor handling of student mental health. So it was not the easiest environment to not have parental support, but also I did not realize that that’s something that was making things more difficult. It was pretty overwhelming.

How has that experience shaped your career path since?

If I had not been first-generation, it’s more likely I would have continued straight into applying to Ph.D. programs. But also after college, I probably would have taken a less winding path than I have taken. And there are some benefits and disadvantages to that. One of the benefits is, when I ended up doing freelance consulting for a while, I dropped out of being in a cohort after spending most of my life in a cohort in an institutional setting. And I got used to being okay with doing things that are not necessarily the typical way to do them. For example, even though the PMP is typically an evening program, during one quarter the programming languages course was taught in parallel with the “normal” morning one. I had the flexibility to take that morning class instead and spend time with the Ph.D. students. One of the professors then invited me to spend time in the Programming Languages & Software Engineering (PLSE) lab, so I again got to interact with Ph.D. students there. It’s put me into a mentality where I think less in terms of a structured path. 

That flexibility is useful, but sometimes it would be nice to have more structure. Another challenge is that if you don’t have this very clear box to show to people, they’re not sure what to make of you. So, for instance, maybe you’re interested in doing research — but people aren’t sure even logistically how that would work. The PMP is not a research program; but at the same time, it’s this great, very broad survey program. So as I’m taking courses in these different areas, I’m thinking about which one sparks the most interest. But one area where the first-gen experience comes in is, I don’t know how to take that and follow through to make an ongoing connection. A lot of students do PMP for professional development, and that’s what it is primarily set up for, but there are also students, like me, who have that intellectual itch and this is the most accessible foot back in the door.

What advice would you give to other first-gen college students?

Something that I think is good advice for undergrads in general is to go to office hours, which as an undergrad, I did not do. Coming back to school and paying for the classes myself — and being really excited about them — face to face time with the professors is so important. Go to office hours even if you are behind on the assignment, or don’t have questions about the assignment, just to listen to what other students are asking. Even if you don’t have anything prepared, some conversation will happen. That has been really helpful for me coming back to the PMP. 

Some people who are first-gen students are very aware of it; it’s part of their identity right from the get-go. But for others, like me, we don’t realize right away how much being first-gen impacts our experience. So keep in mind that you are carrying a lot more weight than other students are. If it seems harder, if it seems things are not coming as easily to you, that’s not surprising. It’s also not your fault, so practice self-compassion.

Nicole Sullivan: Advancing science and sustainability while assisting others in their journey

Nicole Sullivan first began to consider a career in computing-related research as a high school student in Cerritos, California. After enrolling in a computer science course in the 11th grade, she became fascinated with the field’s potential to address environmental challenges such as nature conservation, climate change, agriculture and more. She found further inspiration as a Karsh STEM Scholar and undergraduate researcher at Howard University, an experience she credits with setting her on the path to earning a Ph.D. She followed that path across the country to the UW, where Sullivan is making meaningful contributions to data science and sustainability working alongside professor Magdalena Balazinska in the Allen School’s Ph.D. program. She is also helping to inspire a new generation of researchers by mentoring underrepresented minority students hoping to follow in her footsteps.

What did it mean to you and/or your family to be the first to pursue a bachelor’s degree?

I’m grateful for the opportunity to pursue higher education, which my parents fully support. They were proud of my independence during my undergraduate studies and are ecstatic about my pursuit of a Ph.D.  

What was the most challenging aspect of being a first-gen student, and what was the most rewarding?

While it was overwhelming and challenging to figure out scholarship and college applications on my own, I’ve gained a solid understanding of the process. That has equipped me with valuable insights that enable me to assist others in their journey. 

How have you applied your experience to assist others?

I am currently a graduate mentor for A Vision for Electronic Literacy and Access (AVELA). Before that, while I was at Howard University, I was a National Society of Black Engineers (NSBE) Jr. Mentor and a Microsoft Code Academy (MCA) Lead Learner. As an AVELA mentor, I create and teach original STEM content for Black, Brown, and Indigenous middle and high school students throughout greater Seattle. In NSBE Jr., I supported two teams on their way to the NSBE national robotics competition. Additionally, through MCA, I spent every other weekend teaching programming fundamentals to Black students in kindergarten through 5th grade.

What advice would you give to aspiring first-gen college students?

I highly recommend participating in programs like AVELA and NSBE. Engaging with AVELA, which offers free courses in coding basics, machine learning, hardware, and more, can provide an excellent foundation in various areas of study. Not only will you acquire valuable knowledge, but you’ll also have compelling experiences to highlight in your college application essays and add to your resumes. Furthermore, NSBE extends scholarships to high school students and organizes an annual conference where you might discover field-related opportunities and gain hands-on experience during your high school years. While the NSBE conference isn’t free, consider contacting your high school counselor or local NSBE chapter to explore potential funding options.

What do you hope to do after earning your Ph.D. from the Allen School?

Although I’m not exactly sure what will happen after I graduate, I know I will choose a path either in academia or in industry research. And I intend to continue mentoring underrepresented minority students to foster their enthusiasm for STEM Ph.D. programs and higher education in general.

Learn more about the First Generation College Celebration at UW here.

Read more →