SFI has always prided itself on its ability to bring together top scientists from around the world. Traditionally, they've met in the same room, with catered meals and coffee on tap.
Now, in an effort to help slow the spread of the COVID-19 pandemic, SFI’s faculty, postdocs, and staff are making the most of remote work.
Researchers are holding a weekly teleconference to catch up on each others' science. Called “remotely together” in the initial email invitation from Professor Cris Moore, the Wednesday meetings consist of seminar presentations, informal science talks, and, in Moore's words, “the usual . . . lunchtime conversation.”
Davis Professor Melanie Mitchell delivered the first seminar on March 18, presenting her current research into how we can evaluate an AI’s ability to form abstract concepts. For example, an artificial neural network can accurately identify a picture of a cable bridge, but cannot extend this understanding to the “bridge” of a nose, or a “bridge” between scholarly disciplines. Mitchell presented some classic and state-of-the art problem sets that have been proposed to test this type of abstraction, including one designed by her mentor, Douglas Hofstadter. In classic SFI style, she invited participants to interrupt with questions.
SFI’s IT department set up the video conference, and despite a few audio and visual glitches, participants could speak up throughout the presentation. Interlocutors quickly learned to state their names before speaking, so others could identify them.
“The online format took a little getting used to,” Mitchell says, “but the essential experience of dialogue and camaraderie went amazingly smoothly.”
Toward the end of the presentation, Mitchell returned to the bridge example with a prescient challenge from the field of AI. One explanation for how humans are able to abstract a concept like “bridge” has to do with the way we experience the world as physical, embodied beings. If an AI has never used a plank to cross a stream, or driven the Golden Gate, can it really abstract the ultimate purpose of the object?
Participants unmuted their microphones to applaud, then rallied for an overtime Q&A, approximating the “embodied” experience of lingering in the conference room to continue a lively conversation on a complex problem.