Across the country, schools are wrestling with the difficult choice of whether to reopen, and how to do it with reduced risk. In Kalamazoo, Michigan—not far from one the main sites where Pfizer is frantically manufacturing vaccines—they plan to stay virtual through the end of the school year. In Iowa, a state without a mask mandate, kids can now go back to in-person learning full time. Meanwhile, in a school district in San Mateo County, California, that borders Silicon Valley, there’s no clear decision—and low-income and affluent parents are clashing over what to do.
It’s been a difficult journey. Since March 2020, when most schools closed, districts have been asked to adjust over and over—to new science about how the virus behaves, new policy recommendations, and the different needs of families, kids, teachers, and staff.
Now, as President Biden forges ahead with his promise to reopen most schools within his first 100 days, the debates sound as complicated as ever—and offer a glimpse into many of the difficulties of reopening society at large.
The limits of “guidance”
Schools across the country have looked to the Centers for Disease Control and Prevention for guidance on how to operate in the pandemic. In its latest recommendations, the CDC says a lot of the things we’ve heard all year: that everyone in a school building should wear masks, stay at least six feet apart, and wash their hands frequently. But schools have found that even when guidelines seem relatively straightforward on paper, they are often much harder—or downright impossible—to put into practice.
“There’s a difference between public health mitigation policies when we think them through and when we write them down, and then when we try to implement them,” says Theresa Chapple, an epidemiologist in Washington, DC. “We see that there are barriers at play.”
Chapple points to a recent study by the CDC that looked at elementary schools in Georgia. After just 24 days of in-person learning, the researchers found nine clusters of covid-19 cases that could be linked back to the school. In all, about 45 students and teachers tested positive. How did that happen? Classroom layouts and class sizes meant physical distancing wasn’t possible, so students were less than three feet apart, separated only by plastic dividers. And though students and teachers mostly wore masks, students had to eat lunch in their classrooms.
Researchers also note that teachers and students may have infected each other “during small group instruction sessions in which educators worked in close proximity to students.”
Following the CDC’s best practices might be inherently difficult, but it’s also complicated by the fact that they are just guidelines: states and other jurisdictions make the rules, and those often conflict with what the CDC says to do. Since February 15, Iowa schools have been required to offer fully in-person learning options that some school officials say make distancing impossible. Because the state no longer has a mask mandate, students aren’t required to wear masks in school.
Jurisdictions following all these different policies have one thing in common: although case totals have dipped since their peak in January, the vast majority of the US still has substantial or high community spread. A big takeaway from the CDC’s latest guidance is that high community transmission is linked to increased risk in schools.
“If we are opening schools,” Chapple says, “we are saying that there’s an acceptable amount of spread that we will take in order for children to be educated.”
Meeting different needs
Some schools are trying alternative tactics that they hope will reduce the risks associated with in-person learning.
In Sharon, a Massachusetts town just south of Boston where about 60% of public school students are still learning remotely, pods of students and staff are called down to a central location in their school building twice a week for voluntary covid-19 testing. One by one, children as young as five turn up, sanitize their hands, lower their mask, swab their own nostrils, and place their swab in a single test tube designated for their whole cohort. To make room for everyone, sometimes even the principal’s office becomes a testing site: one person in, one person out. The tubes are then sent to a lab for something called “pooled testing.”
Pooled testing allows a small group of samples to be tested for covid all at once. In Sharon, each tube holds anywhere from 5 to 25 samples. If the test for that small group comes back negative, the whole group is cleared. If it’s positive, each group member is tested until the positive individual is found. Meg Dussault, the district’s acting superintendent, says each pool test costs the school between $5 and $50, and over a third of Sharon Public Schools students and staff participate.
“I’ve seen the benefits of this,” she says “And I believe it’s essential.”
Because schools are funded unequally and largely through taxes, access to resources is a common theme in discussions of school reopening. The state paid for Sharon’s pilot period, but not every district or school has the money or staffing to mount large-scale programs—and Dussault says the district will need to foot the bill for any testing once this program ends in April. It will also need to keep relying on the goodwill of the parent volunteers who wrangle students and swabs for testing each week.
In the seven weeks since pooled testing began, Dussault says, only one batch has come back positive. It’s given her peace of mind.
And even with mitigation measures in place, there are stark demographic differences in opinion on reopening. A recent Pew study found that Black, Asian, and Hispanic adults are more likely to support holding off until teachers have access to vaccines. Those groups are also more likely than white adults to say that the risk of covid-19 transmission “should be given a lot of consideration” when weighing reopening.
Chapple worries that these parents’ concerns will be overlooked, or that funds for remote learning will dwindle because some districts decide to move to in-person learning.
She says: “School districts need to keep in mind that if they’re reopening but a small percentage of their minority students are coming back, what does that look like in terms of equity?”
NASA selects SpaceX’s Starship as the lander to take astronauts to the moon
Surprising selection: Last year, NASA awarded three different groups contracts to further develop their own proposals for lunar landers: $135 million to SpaceX, $253 million to defense company Dynetics (which was working with Sierra Nevada Corporation), and $579 million to a four-company team led by Blue Origin (working with Northrop Grumman, Lockheed Martin, and Draper).
SpaceX didn’t just receive the least amount of money—its proposal also earned the worst technical and management ratings. NASA’s associate administrator (now acting administrator) Steve Jurczyk wrote (pdf) that Starship’s propulsion system was “notably complex and comprised of likewise complex individual subsystems that have yet to be developed, tested, and certified with very little schedule margin to accommodate delays.” The uncertainties were only exacerbated by SpaceX’s notoriously poor track record with meeting deadlines.
What changed: Since then, SpaceX has gone through a number of different flight tests of several full-scale Starship prototypes, including a 10-kilometer high-altitude flight and safe landing in March. (It also exploded a few times.) According to the Washington Post, documents suggest NASA was enamored with Starship’s ability to ferry a lot of cargo to the moon (up to 100 tons), not to mention its $2.9 billion bid for the contract, which was far lower than its rivals’.
“This innovative human landing system will be a hallmark in spaceflight history,” says Lisa Watson-Morgan, NASA’s program manager for the lunar lander system. “We’re confident in NASA’s partnership with SpaceX.”
What this means: For SpaceX’s rivals, it’s a devastating blow—especially to Blue Origin. The company, founded by Jeff Bezos, had unveiled its Blue Moon lander concept in 2019 and has publicly campaigned for NASA to select it for future lunar missions. Blue Moon was arguably the most well-developed of the three proposals when NASA awarded its first round of contracts.
For SpaceX, it’s a big vote of confidence in Starship as a crucial piece of technology for the next generation of space exploration. It comes less than a year after the company’s Crew Dragon vehicle was certified as the only American spacecraft capable of taking NASA astronauts to space. And it seems to confirm that the SpaceX is now NASA’s biggest private partner, supplanting veteran firms like Northrop Grumman and shunting newer ones like Blue Origin further to the sidelines. However, there’s at least one major hurdle: Starship needs to launch using a Super Heavy rocket—a design that SpaceX has yet to fly.
For NASA, the biggest implication is that SpaceX’s vehicles will only continue to play a bigger role for Artemis, the lunar exploration program being touted as the successor to Apollo. Former president Donald Trump’s directive for NASA to return astronauts to the moon by 2024 was never actually going to be realized, but the selection of a single human lander concept suggests NASA may not miss that deadline by much. The first Artemis missions will use Orion, and the long-delayed Space Launch System rocket is expected to be ready soon.
Geoffrey Hinton has a hunch about what’s next for AI
Deep learning set off the latest AI revolution, transforming computer vision and the field as a whole. Hinton believes deep learning should be almost all that’s needed to fully replicate human intelligence.
But despite rapid progress, there are still major challenges. Expose a neural net to an unfamiliar data set or a foreign environment, and it reveals itself to be brittle and inflexible. Self-driving cars and essay-writing language generators impress, but things can go awry. AI visual systems can be easily confused: a coffee mug recognized from the side would be an unknown from above if the system had not been trained on that view; and with the manipulation of a few pixels, a panda can be mistaken for an ostrich, or even a school bus.
GLOM addresses two of the most difficult problems for visual perception systems: understanding a whole scene in terms of objects and their natural parts; and recognizing objects when seen from a new viewpoint.(GLOM’s focus is on vision, but Hinton expects the idea could be applied to language as well.)
An object such as Hinton’s face, for instance, is made up of his lively if dog-tired eyes (too many people asking questions; too little sleep), his mouth and ears, and a prominent nose, all topped by a not-too-untidy tousle of mostly gray. And given his nose, he is easily recognized even on first sight in profile view.
Both of these factors—the part-whole relationship and the viewpoint—are, from Hinton’s perspective, crucial to how humans do vision. “If GLOM ever works,” he says, “it’s going to do perception in a way that’s much more human-like than current neural nets.”
Grouping parts into wholes, however, can be a hard problem for computers, since parts are sometimes ambiguous. A circle could be an eye, or a doughnut, or a wheel. As Hinton explains it, the first generation of AI vision systems tried to recognize objects by relying mostly on the geometry of the part-whole-relationship—the spatial orientation among the parts and between the parts and the whole. The second generation instead relied mostly on deep learning—letting the neural net train on large amounts of data. With GLOM, Hinton combines the best aspects of both approaches.
“There’s a certain intellectual humility that I like about it,” says Gary Marcus, founder and CEO of Robust.AI and a well-known critic of the heavy reliance on deep learning. Marcus admires Hinton’s willingness to challenge something that brought him fame, to admit it’s not quite working. “It’s brave,” he says. “And it’s a great corrective to say, ‘I’m trying to think outside the box.’”
The GLOM architecture
In crafting GLOM, Hinton tried to model some of the mental shortcuts—intuitive strategies, or heuristics—that people use in making sense of the world. “GLOM, and indeed much of Geoff’s work, is about looking at heuristics that people seem to have, building neural nets that could themselves have those heuristics, and then showing that the nets do better at vision as a result,” says Nick Frosst, a computer scientist at a language startup in Toronto who worked with Hinton at Google Brain.
With visual perception, one strategy is to parse parts of an object—such as different facial features—and thereby understand the whole. If you see a certain nose, you might recognize it as part of Hinton’s face; it’s a part-whole hierarchy. To build a better vision system, Hinton says, “I have a strong intuition that we need to use part-whole hierarchies.” Human brains understand this part-whole composition by creating what’s called a “parse tree”—a branching diagram demonstrating the hierarchical relationship between the whole, its parts and subparts. The face itself is at the top of the tree, and the component eyes, nose, ears, and mouth form the branches below.
One of Hinton’s main goals with GLOM is to replicate the parse tree in a neural net—this would distinguish it from neural nets that came before. For technical reasons, it’s hard to do. “It’s difficult because each individual image would be parsed by a person into a unique parse tree, so we would want a neural net to do the same,” says Frosst. “It’s hard to get something with a static architecture—a neural net—to take on a new structure—a parse tree—for each new image it sees.” Hinton has made various attempts. GLOM is a major revision of his previous attempt in 2017, combined with other related advances in the field.
“I’m part of a nose!”
A generalized way of thinking about the GLOM architecture is as follows: The image of interest (say, a photograph of Hinton’s face) is divided into a grid. Each region of the grid is a “location” on the image—one location might contain the iris of an eye, while another might contain the tip of his nose. For each location in the net there are about five layers, or levels. And level by level, the system makes a prediction, with a vector representing the content or information. At a level near the bottom, the vector representing the tip-of-the-nose location might predict: “I’m part of a nose!” And at the next level up, in building a more coherent representation of what it’s seeing, the vector might predict: “I’m part of a face at side-angle view!”
But then the question is, do neighboring vectors at the same level agree? When in agreement, vectors point in the same direction, toward the same conclusion: “Yes, we both belong to the same nose.” Or further up the parse tree. “Yes, we both belong to the same face.”
Seeking consensus about the nature of an object—about what precisely the object is, ultimately—GLOM’s vectors iteratively, location-by-location and layer-upon-layer, average with neighbouring vectors beside, as well as predicted vectors from levels above and below.
However, the net doesn’t “willy-nilly average” with just anything nearby, says Hinton. It averages selectively, with neighboring predictions that display similarities. “This is kind of well-known in America, this is called an echo chamber,” he says. “What you do is you only accept opinions from people who already agree with you; and then what happens is that you get an echo chamber where a whole bunch of people have exactly the same opinion. GLOM actually uses that in a constructive way.” The analogous phenomenon in Hinton’s system is those “islands of agreement.”
“Imagine a bunch of people in a room, shouting slight variations of the same idea,” says Frosst—or imagine those people as vectors pointing in slight variations of the same direction. “They would, after a while, converge on the one idea, and they would all feel it stronger, because they had it confirmed by the other people around them.” That’s how GLOM’s vectors reinforce and amplify their collective predictions about an image.
GLOM uses these islands of agreeing vectors to accomplish the trick of representing a parse tree in a neural net. Whereas some recent neural nets use agreement among vectors for activation, GLOM uses agreement for representation—building up representations of things within the net. For instance, when several vectors agree that they all represent part of the nose, their small cluster of agreement collectively represents the nose in the net’s parse tree for the face. Another smallish cluster of agreeing vectors might represent the mouth in the parse tree; and the big cluster at the top of the tree would represent the emergent conclusion that the image as a whole is Hinton’s face. “The way the parse tree is represented here,” Hinton explains, “is that at the object level you have a big island; the parts of the object are smaller islands; the subparts are even smaller islands, and so on.”
According to Hinton’s long-time friend and collaborator Yoshua Bengio, a computer scientist at the University of Montreal, if GLOM manages to solve the engineering challenge of representing a parse tree in a neural net, it would be a feat—it would be important for making neural nets work properly. “Geoff has produced amazingly powerful intuitions many times in his career, many of which have proven right,” Bengio says. “Hence, I pay attention to them, especially when he feels as strongly about them as he does about GLOM.”
The strength of Hinton’s conviction is rooted not only in the echo chamber analogy, but also in mathematical and biological analogies that inspired and justified some of the design decisions in GLOM’s novel engineering.
“Geoff is a highly unusual thinker in that he is able to draw upon complex mathematical concepts and integrate them with biological constraints to develop theories,” says Sue Becker, a former student of Hinton’s, now a computational cognitive neuroscientist at McMaster University. “Researchers who are more narrowly focused on either the mathematical theory or the neurobiology are much less likely to solve the infinitely compelling puzzle of how both machines and humans might learn and think.”
Turning philosophy into engineering
So far, Hinton’s new idea has been well received, especially in some of the world’s greatest echo chambers. “On Twitter, I got a lot of likes,” he says. And a YouTube tutorial laid claim to the term “MeGLOMania.”
Hinton is the first to admit that at present GLOM is little more than philosophical musing (he spent a year as a philosophy undergrad before switching to experimental psychology). “If an idea sounds good in philosophy, it is good,” he says. “How would you ever have a philosophical idea that just sounds like rubbish, but actually turns out to be true? That wouldn’t pass as a philosophical idea.” Science, by comparison, is “full of things that sound like complete rubbish” but turn out to work remarkably well—for example, neural nets, he says.
GLOM is designed to sound philosophically plausible. But will it work?
Why is it so hard to review the Johnson & Johnson vaccine? Data.
“We’ll never have perfect data, and there will always be uncertainty,” said Grace Lee, a professor at Stanford University and chair of the advisory panel’s Covid-19 Vaccine Safety Technical Subgroup, when the group met on Wednesday. “It’s really, for me, about getting better risk estimates.”
Committee members agreed to reconvene once they’ve had more time to gather and assess data about who might be most at risk of complications, and how that compares with the risk of catching and spreading covid.
All six of the cases reported after the vaccine became widely available occurred in women; one additional case—a man—was reported during clinical trials. All patients were between 18 and 48, and several were treated with the blood thinner heparin, which is typically used for clots but worsened the condition of these patients. The symptoms appear very similar to ones associated with AstraZeneca’s covid vaccine, which many European countries have limited or even stopped using. The active components of both are delivered to cells by adenoviruses that have been modified so they can’t replicate.
But because there are other treatments available that use totally different methods, experts say it is sensible to hold off to see if more information becomes available. The Johnson & Johnson vaccine counts for only 7.5 million of America’s 195 million shots delivered; the Pfizer-BioNTech and Moderna vaccines, which use mRNA rather than adenoviruses, are responsible for the rest.
“The risks and benefits of continuing to administer the J&J vaccine can’t be looked at in isolation,” says Seema Shah, a bioethicist at Lurie Children’s Hospital in Chicago. “If people have alternatives, at least while the FDA is figuring things out, it makes sense to steer people in the direction of those alternatives.”
Resumption of Johnson & Johnson shots may not mean that it becomes available to everybody, however. Ensuring the safety of vaccines is especially important because they’re given to healthy people, rather than treating people who are already sick, and successfully figuring out which groups might see the most benefit—or most harm—could lead to tiered recommendations. Several EU countries, for instance, have said the AstraZeneca vaccine should be given to older people at higher risk of complications from covid, rather than younger people who might be at higher risk of vaccine complications.
“At the end of the day, the critical issue is if I’m a 30-year-old woman and I get this vaccine, how much will that increase my risk of this bad thing?” says Arthur Reingold, chair of California’s Covid-19 Scientific Safety Review Workgroup and a former member of the CDC’s vaccine advisory panel.
A more complicated question is what data the committee will review to make a final decision.
No comprehensive data
Information may be limited because the issue was caught quickly, and because the Johnson & Johnson vaccine is so far being deployed only in the US (the company said it was delaying delivery to European Union countries). But making a determination may also prove difficult because America’s medical data is highly fragmented.
Without a national health-care system, there’s no comprehensive way to assess risks and benefits for different groups that have received the vaccine. There is no routine federal capability to connect patient data with vaccine records. Instead, regulators hope clinicians will hear about the pause and proactively report cases they hadn’t previously connected to vaccinations.
“It might stimulate some clinician to say, ‘Oh my God, Mrs. Jones had that three weeks ago,’” says Reingold. In addition, he says, “there’s still quite a few people who have gotten a dose within the last two weeks, and some of them could develop this rare side effect.”
The voluntary system may seem archaic, but that is how the six cases under review came to the attention of the authorities. They were reported to the CDC through an online database called the Vaccine Adverse Events Reporting System, or VAERS. It is an open website that medical staff, patients, and caregivers can use to notify the government about potential vaccine side effects.
Because the system is so open, and requires opt-in participation, it’s impossible to calculate exact risks using VAERS data. Epidemiologists generally think of it as a place to look for hypotheses that tie vaccines to side effects, rather than a source that can be used to confirm their suspicions.