Recent quotes:

AI and Compute

The total amount of compute, in petaflop/s-days, used to train selected results that are relatively well known, used a lot of compute for their time, and gave enough information to estimate the compute used. A petaflop/s-day (pfs-day) consists of performing 1015 neural net operations per second for one day, or a total of about 1020 operations. The compute-time product serves as a mental convenience, similar to kW-hr for energy. We don’t measure peak theoretical FLOPS of the hardware but instead try to estimate the number of actual operations performed. We count adds and multiplies as separate operations, we count any add or multiply as a single operation regardless of numerical precision (making “FLOP” a slight misnomer), and we ignore ensemble models. Example calculations that went into this graph are provided in this appendix. Doubling time for line of best fit shown is 3.4 months.

An AI startup tries to take better pictures of the heart

Caption Health provided me with unpublished data from a study in which 8 nurses with no previous experience in cardiac ultrasound performed four different types of scans on 240 patients. For assessing patients’ left ventricular size and function, as well as assessment of pericardial effusion, or fluid around the heart, the AI took the same number of usable images. For each, 240 scans were performed, and 237, or 98.8%, were of sufficient quality, according to a panel of five cardiologists. For images of the right ventricle, which is harder to see, the results were a bit worse: 222 images, or 92.5% of them, were of adequate quality. Eric Topol, the director and founder of the Scripps Research Translational Institute, commented that this was still a small number of samples for AI work; Caption Health said it “respectfully disagrees” because the study was prospective. The goal of the study was to show the test was 80% accurate.

Data standards may be wonky, but they will transform health care - STAT

The proposed rule creates a highly promising road map toward the easy exchange of electronic health information that exemplifies a minimalist regulatory approach for creating the standardization and uniformity needed to spark an apps marketplace. It would also create economic and commercial guardrails to promote a level playing field between electronic health record vendors and app developers. These regulations are an essential ingredient for a burgeoning apps market. All six individuals who previously served as the national coordinator of health information technology have endorsed the rule. It has sparked robust conversation: During the public comment period on the proposed rule, nearly 2,000 comments were submitted about interoperability and information blocking. As might be expected, there is pushback from the electronic health record industry on timelines and price controls. The proposed timeline — two years of development— has proven highly realistic, given the successful implementation of SMART on FHIR among the major brands of electronic health records by the Argonaut working group in just one year, and the work of the CARIN alliance to help connect patient apps to the SMART API.

First smart speaker system that uses white noise to monitor infants' breathing - ScienceBlog.com

Detecting breathing in babies has an extra wrinkle: the movement of their chests is so tiny that the smart speaker needs to know exactly where the babies are to be able to “see” them breathing. “The breathing signal is so weak that we can’t just look for a change in the overall signal we get back,” Wang said. “We needed a way to scan the room and pinpoint where the baby is to maximize changes in the white noise signal. Our algorithm takes advantage of the fact that smart speakers have an array of microphones that can be used to focus in the direction of the infant’s chest. It starts listening for changes in a bunch of potential directions, and then continues the search toward the direction that gives the clearest signal.”

Wrist-worn step trackers accurate in predicting patient health outcomes -- ScienceDaily

In the study, researchers conducted a 12-week, blinded, randomized, cross-over trial with 52 patients, a group that included adults with a history of respiratory problems during periods of elevated air pollution. Wrist step counters tracked patient steps for those 12 weeks; and patients also filled out respiratory symptom questionnaires. Researchers found they could effectively estimate a patient's 6MWD results by using step counters, instead of having patients come in a clinical setting to do the 6MWD test. "Instead of having one measurement every few months, you could have weekly measurements, and have information at disease progression at more frequent intervals. This is a significant improvement and enhanced convenience for our patients," said Dr. Blagev. The implications? Using wrist step counters will allow physicians to track how their patients are doing, the progression of the disease, and whether a patient requires an immediate intervention. "Being able to distill step counts into this clinically important metric is a first step in being able to think about how to use step counters in order to better manage health and detect deterioration earlier," Dr. Blagev added.

New chip poised to enable hand-held microwave imaging: Researchers shrink bulky imaging systems down to millimeter sized chip that could be used to see through walls or detect tumors -- ScienceDaily

"Today's practical microwave imagers are bench-top systems that are bulky and expensive," said research team leader Firooz Aflatouni from the University of Pennsylvania, USA. "Our new near-field imager uses optical, rather than electronic, devices to process the microwave signal. This enabled us to make a chip-based imager similar to the optical camera chips in many smartphones." Hand-held near-field microwave imagers would be useful for many applications including high-resolution brain imaging and monitoring heart motion and breathing. Miniaturization of microwave imagers would also benefit applications such as tracking objects in radar systems and low-power, high-speed communication links.

Popular mobile games can be used to detect signs of cognitive decline -- ScienceDaily

Their research put 21 healthy participants through standard paper-based cognitive assessment tests, followed by 10-minute sessions of playing Tetris, Candy Crush Saga and Fruit Ninja over two separate periods, a fortnight apart. The three games selected were chosen because they are easy to learn, engaging for most players and involve intensive interactions using multiple gestures. Using the sensors built into the mobile phones to collect data, the team showed how users interacted with the games and illustrated a clear link between the subjects' touch gestures, or taps and swipes, their rotational gestures and their levels of cognitive performance. The study revealed the participants' ability to perform visuo-spatial and visual search tasks, as well as testing their memory, mental flexibility and attention span.

AI technique detects heart failure from a single heartbeat with 100% accuracy - ScienceBlog.com

Dr Massaro said: “We trained and tested the CNN model on large publicly available ECG datasets featuring subjects with CHF as well as healthy, non-arrhythmic hearts. Our model delivered 100% accuracy: by checking just one heartbeat we are able detect whether or not a person has heart failure. Our model is also one of the first known to be able to identify the ECG’ s morphological features specifically associated to the severity of the condition.”

Cells use sugars to communicate at the molecular level -- ScienceDaily

Using engineered synthetic cells as a model system, lead author Cesar Rodriguez-Emmenegger, a former member of Percec's group, now at Aachen, discovered a way to directly study cell membranes using a method called atomic force microscopy. This approach generates extremely high-resolution scans that reveal shapes and structures at a scale of less than a nanometer, nearly 10,000 times smaller than the width of a human hair. Percec's group then built a model that computes how the structural images relate to the cell's function. The study is the first example of a diffraction-like method that can be done on whole synthetic cells. Using this new method, Percec's group discovered that a lower concentration of sugars on a cell membrane's surface led to increased reactivity with proteins on the membranes of other cells.

Facebook Open Sources Its AI Hardware as It Races Google | WIRED

Big Sur includes eight GPU boards, each loaded with dozens of chips while consuming only about 300 Watts of power. Although GPUs were originally designed to render images for computer games and other highly graphical applications, they’ve proven remarkably adept at deep learning. […]Traditional processors help drive these machines, but big companies like Facebook and Google and Baidu have found that their neural networks are far more efficient if they shift much of the computation onto GPUs. […] After 18 months of development, Big Sur is twice as fast as the previous system Facebook used to train its neural networks.
We’re rushing headlong into the era of cheap cell phones. The peace dividends of the smartphone wars mean you can buy a pretty amazing piece of hardware for what many people spend on lattes each month. That Alcatel has 4G, a quadcore processor, a 13-megapixel camera, and it plays 1080P video. It runs Android 4.2, which isn’t completely current but isn’t totally out of date either, and you can grab one for as little as $80 without a contract. That $129 Moto E ($79 if you get a contract, which you shouldn’t) runs Android 4.4.2, sports a Gorilla Glass screen, has an all-day battery and is even water resistant.
Alice never could quite make out, in thinking it over afterwards, how it was that they began: all she remembers is, that they were running hand in hand, and the Queen went so fast that it was all she could do to keep up with her: and still the Queen kept crying `Faster! Faster!' but Alice felt she could not go faster, thought she had not breath left to say so. The most curious part of the thing was, that the trees and the other things round them never changed their places at all: however fast they went, they never seemed to pass anything. `I wonder if all the things move along with us?' thought poor puzzled Alice. And the Queen seemed to guess her thoughts, for she cried, `Faster! Don't try to talk!' Not that Alice had any idea of doing that. She felt as if she would never be able to talk again, she was getting so much out of breath: and still the Queen cried `Faster! Faster!' and dragged her along. `Are we nearly there?' Alice managed to pant out at last. `Nearly there!' the Queen repeated. `Why, we passed it ten minutes ago! Faster! And they ran on for a time in silence, with the wind whistling in Alice's ears, and almost blowing her hair off her head, she fancied. `Now! Now!' cried the Queen. `Faster! Faster!' And they went so fast that at last they seemed to skim through the air, hardly touching the ground with their feet, till suddenly, just as Alice was getting quite exhausted, they stopped, and she found herself sitting on the ground, breathless and giddy. The Queen propped her up against a tree, and said kindly, `You may rest a little now.' Alice looked round her in great surprise. `Why, I do believe we've been under this tree the whole time! Everything's just as it was!' `Of course it is,' said the Queen, `what would you have it?' `Well, in our country,' said Alice, still panting a little, `you'd generally get to somewhere else -- if you ran very fast for a long time, as we've been doing.' `A slow sort of country!' said the Queen. `Now, here, you see, it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that!'
Unfortunately, since “good design” is defined by the user it’s intended for, it’s not just about creating more, and there is no algorithmic “law” for how to get it. It suffers from the phrase that all technologists and investors hate to hear, which is “… it depends.” Whether we want “more” or “less” doesn’t have a single right answer. An example I like to use is about doing the laundry versus eating a cookie. You always want less laundry, but more cookies. One person’s laundry is another’s cookie. And so on.
By 2017, an amount of data equivalent to all the films ever produced will be transmitted over the internet in a three-minute period, according to Cisco, a manufacturer of communications equipment.Internet traffic today per person is measured in gigabytes, with six gigabytes of information exchanged per human per year. In 2017, that number will have risen to 16. By then, global data will be counted in zettabytes – roughly one trillion gigabytes.