MIT Research News' Journal
[Most Recent Entries]
[Calendar View]
Monday, December 2nd, 2019
| Time |
Event |
| 12:00a |
Helping machines perceive some laws of physics Humans have an early understanding of the laws of physical reality. Infants, for instance, hold expectations for how objects should move and interact with each other, and will show surprise when they do something unexpected, such as disappearing in a sleight-of-hand magic trick.
Now MIT researchers have designed a model that demonstrates an understanding of some basic “intuitive physics” about how objects should behave. The model could be used to help build smarter artificial intelligence and, in turn, provide information to help scientists understand infant cognition.
The model, called ADEPT, observes objects moving around a scene and makes predictions about how the objects should behave, based on their underlying physics. While tracking the objects, the model outputs a signal at each video frame that correlates to a level of “surprise” — the bigger the signal, the greater the surprise. If an object ever dramatically mismatches the model’s predictions — by, say, vanishing or teleporting across a scene — its surprise levels will spike.
In response to videos showing objects moving in physically plausible and implausible ways, the model registered levels of surprise that matched levels reported by humans who had watched the same videos.
“By the time infants are 3 months old, they have some notion that objects don’t wink in and out of existence, and can’t move through each other or teleport,” says first author Kevin A. Smith, a research scientist in the Department of Brain and Cognitive Sciences (BCS) and a member of the Center for Brains, Minds, and Machines (CBMM). “We wanted to capture and formalize that knowledge to build infant cognition into artificial-intelligence agents. We’re now getting near human-like in the way models can pick apart basic implausible or plausible scenes.”
Joining Smith on the paper are co-first authors Lingjie Mei, an undergraduate in the Department of Electrical Engineering and Computer Science, and BCS research scientist Shunyu Yao; Jiajun Wu PhD ’19; CBMM investigator Elizabeth Spelke; Joshua B. Tenenbaum, a professor of computational cognitive science, and researcher in CBMM, BCS, and the Computer Science and Artificial Intelligence Laboratory (CSAIL); and CBMM investigator Tomer D. Ullman PhD ’15.
Mismatched realities
ADEPT relies on two modules: an “inverse graphics” module that captures object representations from raw images, and a “physics engine” that predicts the objects’ future representations from a distribution of possibilities.
Inverse graphics basically extracts information of objects — such as shape, pose, and velocity — from pixel inputs. This module captures frames of video as images and uses inverse graphics to extract this information from objects in the scene. But it doesn’t get bogged down in the details. ADEPT requires only some approximate geometry of each shape to function. In part, this helps the model generalize predictions to new objects, not just those it’s trained on.
“It doesn’t matter if an object is rectangle or circle, or if it’s a truck or a duck. ADEPT just sees there’s an object with some position, moving in a certain way, to make predictions,” Smith says. “Similarly, young infants also don’t seem to care much about some properties like shape when making physical predictions.”
These coarse object descriptions are fed into a physics engine — software that simulates behavior of physical systems, such as rigid or fluidic bodies, and is commonly used for films, video games, and computer graphics. The researchers’ physics engine “pushes the objects forward in time,” Ullman says. This creates a range of predictions, or a “belief distribution,” for what will happen to those objects in the next frame.
Next, the model observes the actual next frame. Once again, it captures the object representations, which it then aligns to one of the predicted object representations from its belief distribution. If the object obeyed the laws of physics, there won’t be much mismatch between the two representations. On the other hand, if the object did something implausible — say, it vanished from behind a wall — there will be a major mismatch.
ADEPT then resamples from its belief distribution and notes a very low probability that the object had simply vanished. If there’s a low enough probability, the model registers great “surprise” as a signal spike. Basically, surprise is inversely proportional to the probability of an event occurring. If the probability is very low, the signal spike is very high.
“If an object goes behind a wall, your physics engine maintains a belief that the object is still behind the wall. If the wall goes down, and nothing is there, there’s a mismatch,” Ullman says. “Then, the model says, ‘There’s an object in my prediction, but I see nothing. The only explanation is that it disappeared, so that’s surprising.’”
Violation of expectations
In development psychology, researchers run “violation of expectations” tests in which infants are shown pairs of videos. One video shows a plausible event, with objects adhering to their expected notions of how the world works. The other video is the same in every way, except objects behave in a way that violates expectations in some way. Researchers will often use these tests to measure how long the infant looks at a scene after an implausible action has occurred. The longer they stare, researchers hypothesize, the more they may be surprised or interested in what just happened.
For their experiments, the researchers created several scenarios based on classical developmental research to examine the model’s core object knowledge. They employed 60 adults to watch 64 videos of known physically plausible and physically implausible scenarios. Objects, for instance, will move behind a wall and, when the wall drops, they’ll still be there or they’ll be gone. The participants rated their surprise at various moments on an increasing scale of 0 to 100. Then, the researchers showed the same videos to the model. Specifically, the scenarios examined the model’s ability to capture notions of permanence (objects do not appear or disappear for no reason), continuity (objects move along connected trajectories), and solidity (objects cannot move through one another).
ADEPT matched humans particularly well on videos where objects moved behind walls and disappeared when the wall was removed. Interestingly, the model also matched surprise levels on videos that humans weren’t surprised by but maybe should have been. For example, in a video where an object moving at a certain speed disappears behind a wall and immediately comes out the other side, the object might have sped up dramatically when it went behind the wall or it might have teleported to the other side. In general, humans and ADEPT were both less certain about whether that event was or wasn’t surprising. The researchers also found traditional neural networks that learn physics from observations — but don’t explicitly represent objects — are far less accurate at differentiating surprising from unsurprising scenes, and their picks for surprising scenes don’t often align with humans.
Next, the researchers plan to delve further into how infants observe and learn about the world, with aims of incorporating any new findings into their model. Studies, for example, show that infants up until a certain age actually aren’t very surprised when objects completely change in some ways — such as if a truck disappears behind a wall, but reemerges as a duck.
“We want to see what else needs to be built in to understand the world more like infants, and formalize what we know about psychology to build better AI agents,” Smith says. | | 10:59a |
New treatment could ease the passage of kidney stones Every year, more than half a million Americans visit the emergency room for kidney stone problems. In most cases, the stones eventually pass out of the body on their own, but the process can be excruciatingly painful.
Researchers at MIT and Massachusetts General Hospital have now devised a potential treatment that could make passing kidney stones faster and less painful. They have identified a combination of two drugs that relax the walls of the ureter — the tube that connects the kidneys to the bladder — and can be delivered directly to the ureter with a catheter-like instrument.
Relaxing the ureter could help stones move through the tube more easily, the researchers say.
“We think this could significantly impact kidney stone disease, which affects millions of people,” says Michael Cima, the David H. Koch Professor of Engineering in MIT’s Department of Materials Science and Engineering, a member of MIT’s Koch Institute for Integrative Cancer Research, and the senior author of the study.
This kind of treatment could also make it easier and less painful to insert stents into the ureter, which is sometimes done after a kidney stone is passed, to prevent the tube from becoming blocked or collapsing.
Christopher Lee, a recent PhD recipient in the Harvard-MIT Division of Health Sciences and Technology, is the lead author of the study, which appears today in Nature Biomedical Engineering.
Local drug delivery
Kidney stones are made from hard crystals that accumulate in the kidneys when there is too much solid waste in the urine and not enough liquid to wash it out. It is estimated that about one in 10 people will have a kidney stone at some point in their lives.
Several years ago, Cima and Brian Eisner, who co-directs the Kidney Stone Program at MGH and is also an author of the paper, began thinking about ways to improve the treatment of kidney stones. While some larger stones require surgery, the usual treatment plan is simply to wait for the stones to pass, which takes an average of 10 days. Patients are given painkillers as well as an oral medication that is meant to help relax the ureter, but studies have offered conflicting evidence on whether this drug actually helps. (There are no FDA-approved oral therapies for kidney stones and ureteral dilation.)
Cima and Eisner thought that delivering a muscle relaxant directly to the ureter might offer a better alternative. Most of the pain from passing a kidney stone arises from cramps and inflammation in the ureter as the stones pass through the narrow tube, so relaxing the muscles surrounding the tube could help ease this passage.
Around this time, Lee, then a new student in MIT’s Health Sciences and Technology program, met with Cima to discuss possible thesis projects and became interested in pursuing a kidney stone treatment.
“If you look at how kidney stones are treated today, it hasn’t really changed since about 1980, and there’s a pretty substantial amount of evidence that the drugs given don’t work very well,” Lee says. “The volume of how many people this could potentially help is really exciting.”
The researchers first set out to identify drugs that might work well when delivered directly to the ureter. They selected 18 drugs used to treat conditions such as high blood pressure or glaucoma and exposed them to human ureteral cells grown in a lab dish, where they could measure how much the drugs relaxed the smooth muscle cells. They hypothesized that if they delivered such drugs directly to the ureter, they could get a much bigger relaxation effect than by delivering such drugs orally, while minimizing possible harm to the rest of the body.
“We found several drugs that had the effect that we expected, and in every case we found that the concentrations required to be effective were more than would be safe if given systemically,” Cima says.
Next, the researchers used intensive computational processing to individually analyze the relaxation responses of nearly 1 billion cells after drug exposure. They identified two drugs that worked especially well, and found that they worked even better when given together. One of these is nifedipine, a calcium channel blocker used to treat high blood pressure, and the other is a type of drug known as a ROCK (rho kinase) inhibitor, which is used to treat glaucoma.
The researchers tested various doses of this combination of drugs in ureters removed from pigs, and showed that they could dramatically reduce the frequency and length of contractions of the ureter. Tests in live pigs also showed that the treatment nearly eliminated ureteral contractions.
For these experiments, the researchers delivered the drugs using a cystoscope, which is very similar to a catheter but has a small fiber optic channel that can connect to a camera or lens. They found that with this type of delivery, the drugs were not detectable in the animals’ bloodstream, suggesting that the drugs remained in the lining of the ureter and did not go elsewhere in the body, which would lessen the risk of potential side effects.
Ureteral relaxation
More studies are needed to determine how long the muscle relaxing effect lasts and how much relaxation would be needed to expedite stone passage, the researchers say. They are now launching a startup company, Fluidity Medicine, to continue developing the technology for possible testing in human patients.
In addition to treating kidney stones, this approach could also be useful for relaxing the ureter to help doctors insert a ureteral stent. It could also help when placing any other kind of instrument, such as an endoscope, in the ureter.
“The platform pairs drug delivery to the ureter. We are eager to first target muscle relaxation, and as offshoots of that, we have kidney stones, ureteral stents, and endoscopic surgery,” Lee says. “We have a bunch of other urological indications that would go through different developmental pathways but can all be hit and all have meaningful patient populations.”
The research was funded by the MIT Institute of Medical Engineering and Science Broshy Fellowship, the MIT Deshpande Center for Technological Innovation, the Koch Institute Support (core) Grant from the National Cancer Institute, and the National Institutes of Health. | | 2:59p |
A new way to control microbial metabolism Microbes can be engineered to produce a variety of useful compounds, including plastics, biofuels, and pharmaceuticals. However, in many cases, these products compete with the metabolic pathways that the cells need to fuel themselves and grow.
To help optimize cells’ ability to produce desired compounds but also maintain their own growth, MIT chemical engineers have devised a way to induce bacteria to switch between different metabolic pathways at different times. These switches are programmed into the cells and are triggered by changes in population density, with no need for human intervention.
“What we’re hoping is that this would allow more precise regulation of metabolism, to allow us to get higher productivity, but in a way where we minimize the number of interventions,” says Kristala Prather, the Arthur D. Little Professor of Chemical Engineering and the senior author of the study.
This kind of switching allowed the researchers to boost the microbial yields of two different products by up to tenfold.
MIT graduate student Christina Dinh is the lead author of the paper, which appears in the Proceedings of the National Academy of Sciences this week.
Double switch
To make microbes synthesize useful compounds that they don’t normally produce, engineers insert genes for enzymes involved in the metabolic pathway — a chain of reactions that generate a specific product. This approach is now used to produce many complex products, such as pharmaceuticals and biofuels.
In some cases, intermediates produced during these reactions are also part of metabolic pathways that already exist in the cells. When cells divert these intermediates out of the engineered pathway, it lowers the overall yield of the end product.
Using a concept called dynamic metabolic engineering, Prather has previously built switches that help cells maintain the balance between their own metabolic needs and the pathway that produces the desired product. Her idea was to program the cells to autonomously switch between pathways, without the need for any intervention by the person operating the fermenter where the reactions take place.
In a study published in 2017, Prather’s lab used this approach to program E. coli to produce glucaric acid, a precursor to products such as nylons and detergents. The researchers’ strategy was based on quorum sensing, a phenomenon that bacterial cells normally use to communicate with each other. Each species of bacteria secretes particular molecules that help them sense nearby microbes and influence each other’s behavior.
The MIT team engineered their E. coli cells to secrete a quorum sensing molecule called AHL. When AHL concentrations reach a certain level, the cells shut off an enzyme that diverts a glucaric acid precursor into one of the cells’ own metabolic pathways. This allows the cells to grow and divide normally until the population is large enough to start producing large quantities of the desired product.
“That paper was the first to demonstrate that we could do autonomous control,” Prather says. “We could start the cultures going, and the cells would then sense when the time was right to make a change.”
In the new PNAS paper, Prather and Dinh set out to engineer multiple switching points into their cells, giving them a greater degree of control over the production process. To achieve that, they used two quorum sensing systems from two different species of bacteria. They incorporated these systems into E. coli that were engineered to produce a compound called naringenin, a flavonoid that is naturally found in citrus fruits and has a variety of beneficial health effects.
Using these quorum sensing systems, the researchers engineered two switching points into the cells. One switch was designed to prevent bacteria from diverting a naringenin precursor called malonyl-CoA into the cells’ own metabolic pathways. At the other switching point, the researchers delayed production of an enzyme in their engineered pathway, to avoid accumulating a precursor that normally inhibits the naringenin pathway if too much of the precursor accumulates.
“Since we took components from two different quorum sensing systems, and the regulator proteins are unique between the two systems, we can shift the switching time of each of the circuits independently,” Dinh says.
The researchers created hundreds of E. coli variants that perform these two switches at different population densities, allowing them to identify which one was the most productive. The best-performing strain showed a tenfold increase in naringenin yield over strains that didn’t have these control switches built in.
“The paper addresses an important problem in the area of regulating metabolic pathways to balance cellular growth versus the production of chemicals,” says Radhakrishnan Mahadevan, a professor of chemical engineering at the University of Toronto, who was not involved in the research. “Previously, the circuits primarily focused on turning off genes related to growth, whereas in this contribution they provide the flexibility to downregulate and upregulate specific genes in response to a trigger. This advance should provide more flexible control of metabolic pathways and will be valuable to optimize bioprocesses to improve their economic viability.”
More complex pathways
The researchers also demonstrated that the multiple-switch approach could be used to double E. coli production of salicylic acid, a building block of many drugs. This process could also help improve yields for any other type of product where the cells have to balance between using intermediates for product formation or their own growth, Prather says. The researchers have not yet demonstrated that their method works on an industrial scale, but they are working on expanding the approach to more complex pathways and hope to test it at a larger scale in the future.
“We think it certainly has broader applicability,” Prather says. “The process is very robust because it doesn’t require someone to be present at a particular point in time to add something or make any sort of adjustment to the process, but rather allows the cells to be keeping track internally of when it’s time to make a shift.”
The research was funded by the National Science Foundation. |
|