Wednesday, July 3, 2019
Tuesday, July 2, 2019
The first AI universe sim is fast and accurate—and its creators don't know how it works
by Thomas Sumner, Simons
Foundation
JUNE 26, 2019
For the first time,
astrophysicists have used artificial intelligence techniques to generate
complex 3-D simulations of the universe. The results are so fast, accurate and
robust that even the creators aren't sure how it all works.
"We can run these
simulations in a few milliseconds, while other 'fast' simulations take a couple
of minutes," says study co-author Shirley Ho, a group leader at the
Flatiron Institute's Center for Computational Astrophysics in New York City and
an adjunct professor at Carnegie Mellon University. "Not only that, but
we're much more accurate."
The speed and accuracy of the
project, called the Deep Density Displacement Model, or D3M for short, wasn't
the biggest surprise to the researchers. The real shock was that D3M could
accurately simulate how the universe would look if certain parameters were
tweaked—such as how much of the cosmos is dark matter—even though the model had
never received any training data where those parameters varied.
"It's like teaching image
recognition software with lots of pictures of cats and dogs, but then it's able
to recognize elephants," Ho explains. "Nobody knows how it does this,
and it's a great mystery to be solved."
Ho and her colleagues present
D3M June 24 in the Proceedings of the National Academy of Sciences. The
study was led by Siyu He, a Flatiron Institute research analyst.
Ho and He worked in
collaboration with Yin Li of the Berkeley Center for Cosmological Physics at
the University of California, Berkeley, and the Kavli Institute for the Physics
and Mathematics of the Universe near Tokyo; Yu Feng of the Berkeley Center for
Cosmological Physics; Wei Chen of the Flatiron Institute; Siamak Ravanbakhsh of
the University of British Columbia in Vancouver; and Barnabás Póczos of
Carnegie Mellon University.
Computer simulations like
those made by D3M have become essential to theoretical astrophysics. Scientists
want to know how the cosmos might evolve under various scenarios, such as if
the dark energy pulling the universe apart varied over time. Such studies
require running thousands of simulations, making a lightning-fast and highly
accurate computer model one of the major objectives of modern astrophysics.
D3M models how gravity shapes
the universe. The researchers opted to focus on gravity alone because it is by
far the most important force when it comes to the large-scale evolution of the
cosmos.
The most accurate universe
simulations calculate how gravity shifts each of billions of individual
particles over the entire age of the universe. That level of accuracy takes
time, requiring around 300 computation hours for one simulation. Faster methods can
finish the same simulations in about two minutes, but the shortcuts required
result in lower accuracy.
Ho, He and their colleagues
honed the deep neural
network that powers D3M by feeding it 8,000 different simulations from
one of the highest-accuracy models available. Neural networks take training
data and run calculations on the information; researchers then compare the
resulting outcome with the expected outcome. With further training, neural networks adapt
over time to yield faster and more accurate results.
After training D3M, the
researchers ran simulations of a box-shaped universe 600 million
light-years across and compared the results to those of the slow and fast
models. Whereas the slow-but-accurate approach took hundreds of hours of
computation time per simulation and the existing fast method took a couple of
minutes, D3M could complete a simulation in just 30 milliseconds.
D3M also churned out accurate
results. When compared with the high-accuracy model, D3M had a relative error
of 2.8 percent. Using the same comparison, the existing fast model had a
relative error of 9.3 percent.
D3M's remarkable ability to
handle parameter variations not found in its training data makes it an
especially useful and flexible tool, Ho says. In addition to modeling other
forces, such as hydrodynamics, Ho's team hopes to learn more about how the
model works under the hood. Doing so could yield benefits for the advancement
of artificial intelligence and machine learning, Ho says.
"We can be an interesting
playground for a machine learner to use to see why this model extrapolates so
well, why it extrapolates to elephants instead of just recognizing cats and
dogs," she says. "It's a two-way street between science and deep
learning."
Discovery of a “Holy Grail” with the invention of universal computer memory
Thread starter P4-630
Start date Jun
25, 2019
A new type of computer memory
which could solve the digital technology energy crisis has been invented and
patented by Lancaster scientists.
The electronic memory device –
described in research published in Scientific
Reports - promises to transform daily life with its ultra-low energy
consumption.
In the home, energy savings
from efficient lighting and appliances have been completely wiped out by
increased use of computers and gadgets, and by 2025 a ‘tsunami of data’ is
expected to consume a fifth of global electricity.
But this new device would
immediately reduce peak power consumption in data centres by a fifth.
It would also allow, for
example, computers which do not need to boot up and could instantaneously and
imperceptibly go into an energy-saving sleep mode – even between key
stokes.
The device is the realisation
of the search for a “Universal Memory” which has preoccupied scientists and
engineers for decades.
Physics Professor Manus Hayne
of Lancaster University said: “Universal Memory, which has robustly stored data
that is easily changed, is widely considered to be unfeasible, or even
impossible, but this device demonstrates its contradictory properties.”
A US patent has been awarded
for the electronic memory device with another patent pending, while several
companies have expressed an interest or are actively involved in the
research.
The inventors of the device
used quantum mechanics to solve the dilemma of choosing between stable,
long-term data storage and low-energy writing and erasing.
The device could replace the
$100bn market for Dynamic Random Access Memory (DRAM), which is the ‘working
memory’ of computers, as well as the long-term memory in flash drives.
While writing data to DRAM is
fast and low-energy, the data is volatile and must be continuously ‘refreshed’
to avoid it being lost: this is clearly inconvenient and inefficient. Flash
stores data robustly, but writing and erasing is slow, energy intensive
and deteriorates it, making it unsuitable for working memory.
Professor Hayne said: “The
ideal is to combine the advantages of both without their drawbacks, and this is
what we have demonstrated. Our device has an intrinsic data storage time that
is predicted to exceed the age of the Universe, yet it can record or delete
data using 100 times less energy than DRAM.”
Research: Researchers grow active mini-brain-networks —
July 2, 2019
Cerebral organoids are
artificially grown, 3D tissue cultures that resemble the human brain. Now,
researchers from Japan report functional neural networks derived from these
organoids in a study publishing June 27 in the journal Stem Cell Reports.
Although the organoids aren’t actually “thinking,” the researchers’ new tool —
which detects neural activity using organoids — could provide a method for
understanding human brain function.
“Because they can mimic
cerebral development, cerebral organoids can be used as a substitute for the
human brain to study complex developmental and neurological disorders,” says
corresponding author Jun Takahashi, a professor at Kyoto University.
However, these studies are
challenging, because current cerebral organoids lack desirable supporting
structures, such as blood vessels and surrounding tissues, Takahashi says.
Since researchers have a limited ability to assess the organoids’ neural
activities, it has also been difficult to comprehensively evaluate the function
of neuronal networks.
“In our study, we created a
new functional analysis tool to assess the comprehensive dynamic change of
network activity in a detected field, which reflected the activities of over
1,000 cells,” says first and co-corresponding author Hideya Sakaguchi, a
postdoctoral fellow at Kyoto University (currently at Salk Institute). “The
exciting thing about this study is that we were able to detect dynamic changes
in the calcium ion activity and visualize comprehensive cell activities.”
To generate the organoids,
Takahashi, Sakaguchi, and their team created a ball of pluripotent stem cells
that have the potential to differentiate into various body tissues. Then, they
placed the cells into a dish filled with culture medium that mimicked the
environment necessary for cerebral development. Using the organoids, the team
successfully visualized synchronized and non-synchronized activities in
networks and connections between individual neurons. The synchronized neural
activity can be the basis for various brain functions, including memory.
“We believe that our work
introduces the possibility of a broad assessment of human cell-derived neural
activity,” Sakaguchi says. The method could help researchers understand
processes by which information is encoded in the brain through the activity of
specific cell populations, as well as the fundamental mechanisms underlying
psychiatric diseases, he says.
While cerebral organoids
provide a means for studying the human brain, ethical concerns have been
previously raised regarding the neural function of cerebral organoids.
“Because cerebral organoids
mimic the developmental process, a concern is that they also have mental
activities such as consciousness in the future,” Sakaguchi says. “Some people
have referenced the famous ‘brains in a vat’ thought experiment proposed by
Hilary Putnam, that brains placed in a vat of life-sustaining liquid with
connection to a computer may have the same consciousness as human beings.”
However, Takahashi and
Sakaguchi believe that cerebral organoids are unlikely to develop consciousness
because they lack input from their surrounding environments.
“Consciousness requires
subjective experience, and cerebral organoids without sensory tissues will not
have sensory input and motor output,” Sakaguchi says. “However, if cerebral
organoids with an input and output system develop consciousness requiring moral
consideration, the basic and applied research of these cerebral organoids will
become a tremendous ethical challenge.”
In the future, applied
organoid research will likely explore three main areas — drug discovery,
modelling neuropsychiatric disorders, and regenerative medicine, Takahashi
says.
“Cerebral organoids can bring
great advances to pharmacological companies by replacing traditional animal
models and can also be used to model untreatable neural diseases,” he says.
“Using our method, it will be possible to analyze cell activity patterns in
brain functions to further explore these areas.”
Source:
Cell Press.
How you and your friends can play a video game together using only your minds
July 1, 2019
UW News
Telepathic communication might
be one step closer to reality thanks to new research from the University of
Washington. A team created a method that allows three people to work together
to solve a problem using only their minds.
In BrainNet, three people play
a Tetris-like game using a brain-to-brain interface. This is the first
demonstration of two things: a brain-to-brain network of more than two people,
and a person being able to both receive and send information to others using
only their brain. The team published its results April
16 in the Nature journal Scientific
Reports, though this research previously attracted media attention after
the researchers posted it September
to the preprint site arXiv.
“Humans are social beings who
communicate with each other to cooperate and solve problems that none of us can
solve on our own,” said corresponding author Rajesh Rao, the CJ
and Elizabeth Hwang professor in the UW’s Paul G. Allen School of Computer
Science & Engineering and a co-director of the Center for Neurotechnology. “We wanted to
know if a group of people could collaborate using only their brains. That’s how
we came up with the idea of BrainNet: where two people help a third person
solve a task.”
As in Tetris, the game shows a
block at the top of the screen and a line that needs to be completed at the
bottom. Two people, the Senders, can see both the block and the line but can’t
control the game. The third person, the Receiver, can see only the block but
can tell the game whether to rotate the block to successfully complete the
line. Each Sender decides whether the block needs to be rotated and then passes
that information from their brain, through the internet and to the brain of the
Receiver. Then the Receiver processes that information and sends a command — to
rotate or not rotate the block — to the game directly from their brain,
hopefully completing and clearing the line.
The team asked five groups of
participants to play 16 rounds of the game. For each group, all three
participants were in different rooms and couldn’t see, hear or speak to one
another.
The Senders each could see the
game displayed on a computer screen. The screen also showed the word “Yes” on
one side and the word “No” on the other side. Beneath the “Yes” option, an LED
flashed 17 times per second. Beneath the “No” option, an LED flashed 15 times a
second.
“Once the Sender makes a
decision about whether to rotate the block, they send ‘Yes’ or ‘No’ to the
Receiver’s brain by concentrating on the corresponding light,” said first
author Linxing
Preston Jiang, a student in the Allen School’s combined bachelor’s/master’s
degree program.
The Senders wore
electroencephalography caps that picked up electrical activity in their brains.
The lights’ different flashing patterns trigger unique types of activity in the
brain, which the caps can pick up. So, as the Senders stared at the light for their
corresponding selection, the cap picked up those signals, and the computer
provided real-time feedback by displaying a cursor on the screen that moved
toward their desired choice. The selections were then translated into a “Yes”
or “No” answer that could be sent over the internet to the Receiver.
“To deliver the message to the
Receiver, we used a cable that ends with a wand that looks like a tiny racket
behind the Receiver’s head. This coil stimulates the part of the brain that
translates signals from the eyes,” said co-author Andrea
Stocco, a UW assistant professor in the Department of Psychology and the
Institute for Learning & Brain Sciences, or I-LABS. “We essentially ‘trick’
the neurons in the back of the brain to spread around the message that they
have received signals from the eyes. Then participants have the sensation that
bright arcs or objects suddenly appear in front of their eyes.”
If the answer was, “Yes,
rotate the block,” then the Receiver would see the bright flash. If the answer
was “No,” then the Receiver wouldn’t see anything. The Receiver received input
from both Senders before making a decision about whether to rotate the block.
Because the Receiver also wore an electroencephalography cap, they used the
same method as the Senders to select yes or no.
The Senders got a chance to
review the Receiver’s decision and send corrections if they disagreed. Then,
once the Receiver sent a second decision, everyone in the group found out if
they cleared the line. On average, each group successfully cleared the line 81%
of the time, or for 13 out of 16 trials.
The researchers wanted to know
if the Receiver would learn over time to trust one Sender over the other based
on their reliability. The team purposely picked one of the Senders to be a “bad
Sender” and flipped their responses in 10 out of the 16 trials — so that a
“Yes, rotate the block” suggestion would be given to the Receiver as “No, don’t
rotate the block,” and vice versa. Over time, the Receiver switched from being
relatively neutral about both Senders to strongly preferring the information
from the “good Sender.”
The team hopes that these
results pave the way for future brain-to-brain interfaces that allow people to
collaborate to solve tough problems that one brain alone couldn’t solve. The
researchers also believe this is an appropriate time to start to have a larger
conversation about the ethics of this kind of brain augmentation research and
developing protocols to ensure that people’s privacy is respected as the
technology improves. The group is working with the Neuroethics team at
the Center for Neurotechnology to address these types of issues.
“But for now, this is just a
baby step. Our equipment is still expensive and very bulky and the task is a
game,” Rao said. “We’re in the ‘Kitty Hawk’ days of brain interface
technologies: We’re just getting off the ground.”
See a related story from NPR.
Subscribe to:
Comments (Atom)