The College of Washington’s Paul G. Allen Faculty of Pc Science & Engineering is reframing what it means for its analysis to vary the world.
In unveiling six “Grand Challenges” at its annual Analysis Showcase and Open Home in Seattle on Wednesday, the Allen Faculty’s leaders described a blueprint for expertise that protects privateness, helps psychological well being, broadens accessibility, earns public belief, and sustains individuals and the planet.
The thought is to “set up ourselves into some extra particular grand challenges that we are able to sort out collectively to have a good larger influence,” mentioned Magdalena Balazinska, director of the Allen Faculty and a UW laptop science professor, opening the varsity’s annual Analysis Showcase and Open Home.
Listed below are the six grand challenges:
- Anticipate and deal with safety, privateness, and security points as tech permeates society.
- Make high-quality cognitive and psychological well being help out there to all.
- Design expertise to be accessible at its inception — not as an add-on.
- Design AI in a method that’s clear and equally useful to all.
- Construct techniques that may be trusted to do precisely what we would like them to do, each time.
- Create applied sciences that maintain individuals and the planet.
Balazinska defined that the listing attracts on the strengths and pursuits of its school, who now quantity greater than 90, together with 74 on the tenure observe.
With whole enrollment of about 2,900 college students, final 12 months the Allen Faculty graduated greater than 600 undergrads, 150 grasp’s college students, and 50 Ph.D. college students.
The Allen Faculty has grown so massive that subfields like techniques and NLP (pure language processing) danger changing into remoted “mini departments,” mentioned Shwetak Patel, a College of Washington laptop science professor. The Grand Challenges initiative emerged as a bottom-up effort to reconnect these teams round shared, human-centered issues.
Patel mentioned the initiative additionally encourages collaborations on campus past the pc science faculty, citing examples like fetal coronary heart charge monitoring with UW Medication.
A serial entrepreneur and 2011 MacArthur Fellow, Patel recalled that when he joined UW 18 years in the past, his utilized and entrepreneurial focus was seen as unconventional. Now it’s central to the varsity’s course. The grand challenges initiative is “music to my ears,” Patel mentioned.
In tackling these challenges, the Allen Faculty has a novel benefit towards many different laptop science colleges. Eighteen school members at present maintain what’s generally known as “concurrent engagements” — formally splitting time between the Allen Faculty and firms and organizations similar to Google, Meta, Microsoft, and the Allen Institute for AI (Ai2).
This can be a “superpower” for the Allen Faculty, mentioned Patel, who has a concurrent engagement at Google. These preparations, he defined, give school and college students entry to knowledge, computing assets, and real-world challenges by working instantly with firms growing essentially the most superior AI techniques.
“A whole lot of the issues we’re attempting to resolve, you can’t resolve them simply on the college,” Patel mentioned, pointing to examples similar to open-source basis fashions and AI for mental-health analysis that rely on large-scale assets unavailable in academia alone.
These roles may also stretch professors skinny. “When any individual’s break up, there’s solely a lot psychological power you may put into the college,” Patel mentioned. Lots of these school members train only one or two programs a 12 months, requiring the varsity to rely extra on lecturers and educating school.
Nonetheless, he mentioned, the advantages outweigh the prices. “I’d fairly have 50% of any individual than 0% of any individual, and we’ll make it work,” he mentioned. “That’s been our technique.”
On the night poster session, graduate college students crammed the rooms to showcase their newest initiatives — together with new advances in synthetic intelligence for speech, language, and accessibility.
DopFone: Doppler-based fetal coronary heart charge monitoring utilizing commodity smartphones
DopFone transforms telephones into fetal coronary heart charge screens. It makes use of the telephone speaker to transmit a steady sine wave and makes use of the microphone to report the reflections. It then processes the audio recordings to estimate fetal coronary heart charge. It goals to be an alternative choice to doppler ultrasounds that require educated employees, which aren’t sensible for frequent distant use.
“The key influence could be within the rural, distant and low-resource settings the place entry to such maternity care is much less — additionally known as maternity care deserts,” mentioned Poojita Garg, a second-year PhD pupil.
CourseSLM: A Chatbot Device for Supporting Instructors and Classroom Studying
This tradition-built chatbot is designed to assist college students keep targeted and construct actual understanding fairly than counting on fast shortcuts. The system makes use of built-in guardrails to maintain learners on activity and counter the distractions and over-dependence that may include common massive language fashions.
Working domestically on faculty units, the chatbot helps shield pupil knowledge and ensures entry even with out Wi-Fi.
“We’re targeted on ensuring college students have entry to expertise, and know the way to use it correctly and safely,” mentioned Marquiese Garrett, a sophomore on the UW.
Environment friendly serving of SpeechLMs with VoxServe
VoxServe makes speech-language fashions run extra effectively. It makes use of a standardized abstraction layer and interface that permits many alternative fashions to run by a single system. Its key innovation is a customized scheduling algorithm that optimizes efficiency relying on the use case.
The strategy makes speech-based AI techniques quicker, cheaper, and simpler to deploy, paving the way in which for real-time voice assistants and different next-gen speech purposes.
“I believed it could be useful if we are able to present this type of open-source system that folks can use,” mentioned Keisuke Kamahori, third-year Ph.D. pupil on the Allen Faculty.
ConvFill: Mannequin collaboration for responsive conversational voice brokers
ConvFill is a light-weight conversational mannequin designed to cut back the delay in voice-based massive language fashions. The system responds shortly with quick, preliminary solutions, then fills in additional detailed data as bigger fashions full their processing.
By combining small and enormous fashions on this method, ConvFill delivers quicker responses whereas conserving tokens and enhancing effectivity — an vital step towards extra pure, low-latency conversational AI.
“That is an thrilling method to consider how we are able to mix techniques collectively to get the very best of each worlds,” mentioned Zachary Englhardt, a third-year Ph.D. pupil. “It’s an thrilling method to take a look at issues.”
ConsumerBench: Benchmarking generative AI on end-user units
Working generative AI domestically — on laptops, telephones, or different private {hardware} — introduces new system-level challenges in equity, effectivity, and scheduling.
ConsumerBench is a benchmarking framework that assessments how effectively generative AI purposes carry out on client {hardware} when a number of AI fashions run on the similar time. The open-source instrument helps researchers determine bottlenecks and enhance efficiency on client units.
There are an a variety of benefits to operating fashions domestically: “There are privateness functions — a consumer can ask for questions associated to e mail or personal content material, they usually can do it effectively and precisely,” mentioned Yile Gu, a third-year Ph.D. pupil on the Allen Faculty.
Designing Chatbots for Delicate Well being Contexts: Classes from Contraceptive Care in Kenyan Pharmacies
A venture aimed toward enhancing contraceptive entry and steering for adolescent women and younger girls in Kenya by integrating low-fidelity chatbots into healthcare settings. The aim is to grasp how chatbots can help personal, knowledgeable conversations and work successfully inside pharmacies.
“The gas behind this entire venture is that my staff is actually eager about enhancing well being outcomes for weak populations,” mentioned Lisa Orii, a fifth-year Ph.D. pupil.
See extra concerning the analysis showcase here.
