A few weeks ago, I downloaded an application, one of the over two million in the App Store that supports my phone. But this particular download wasn’t yet another camera filter, or another effort in my attempt to find the perfect hiking companion. It was a program called “iNaturalist” that works to help identify the plants and animals around us.
I took a picture of a large insect pulled from a pool behind a family home — iNaturalist instantly identified it as a a American Giant Water bug. I took a picture of a frog screaming loudly from a perch high on a wall. INaturalist tagged it as a Grey Tree Frog. Over in the garden some long tendrils and a cluster of long, arrowhead-shaped leaves were outed as the first steps of an invasion from Field Bindweed.
All three of these pictures were not staged or professional. They were shots taken from my phone—a bug on the lip of a plastic cup, a frog crouched down on a board, a few leaves in dappled shade. And in every case, I didn’t provide the app a single prompt for what I was looking at — not even whether what I was looking at was a plant or an animal. Since then, iNaturalist has quickly and accurately identified turtles, birds, bugs, trees, with no input from me other than pointing the camera. That little white flower I don’t know the name of? Common Star-of-Bethlehem. That rather nondescript moth turns out to be a Willow Beauty.
The way the app pulls the item of interest from the background for evaluation is a bit of edge-detecting software magic, but the way it got so smart about hammering out those IDs is easier to explain — lots and lots of people. About 70,000 people a day add more than 10,000 new images every day and also go back to look at other people’s images and share their own identifications. Behind the scenes, the app and it’s web counterpart use these feedback to hone identification skills.
At one point in the last couple of weeks … I was a bad use. I uploaded an image that was intentionally nondescript, a small brown bird, and added an ID of my own. That was intentionally wrong. However, within a matter of minutes, both the app and multiple users had appeared to drown out by mis-identification with correct results. They kept piling on—and kept being right, despite me seeding this little experiment with an incorrect response. Which seems to indicate that iNaturalist isn’t just an amazing app, it’s an app that does a good job of capturing the combined ID skills of an amazing and dedicated community of users.
In a lot of ways, the ID skills of iNaturalist are an extension of projects like this one at Zooniverse where participants are asked to identify creatures imaged by trap cameras in a New York forest. Those IDs don’t just go toward answering the immediate question of “What’s in this picture?” They also build up a database that can be used by software to better make these identifications in an automated fashion. And as the first article today reveals, programs are getting very, very good at this game ...
Computer Sciences
Using AI to identify real animals.
A team from multiple US universities, lead by researchers from the University of Wyoming, applied deep learning techniques to trap camera images in an effort to put Zooniverse volunteers like me out of a that’s-not-a-job. And they were pretty darn successful.
Here, we demonstrate that a cutting-edge type of artificial intelligence called deep neural networks can automatically extract such invaluable information. For example, we show deep learning can automate animal identification for 99.3% of the 3.2 million-image Snapshot Serengeti dataset while performing at the same 96.6% accuracy of crowdsourced teams of human volunteers.
Snapshot Serengeti is in fact a Zooniverse Project, and one on which I’ve contributed a few identifications of hyena, honey badgers, and vervet monkeys. Identifying these images is not just an enjoyable pass time, but provides some genuine insight into the conditions in the area and the frequency in which various animals can be found. And, of course, it was clear from the outset that a big part of the purpose of carrying out these IDs was not just to spot a cheetah or caracal, but to train computers to do that task.
The results of that training … turn out to be quite good. Driving off the crowd-sourced datatabase provided by human volunteers, it’s possible not to automate the analysis — even of new images not in the original set — while achieving an accuracy that matches that of humans. So, job well done to both volunteers and programmers. And personally, I’m going to keep going to the various Snapshot projects and making IDs, even if the computers don’t really need me any more. Because doing this sort of identification is challenging enough that it’s not just computers whose skills are improved by the practice. Besides, where else am I going to see an Aardwolf?
Medical Science
Speeding up stem cell production, without losing flexibility.
It’s taken longer than many originally anticipated, but stem cell use is showing up increasingly varied and increasingly power therapies. One problem holding stem cells back from their full potential has been simply producing them in quantity. Previous techniques for growing human stem cells in bioreactors either tended to produce at a slow rate, or forced stem cells to give up some of their pluripotent do-it-all abilities in exchange for faster reproduction. But a team from University of Toronto has developed a new way of providing both rapid production of a large quantity of stem cells, without sacrificing the flexibility of those cells.
High-density suspension cultures enable process intensification, cost reduction, and more efficient manufacturing. This work advances cell-state engineering as a valuable tool to overcome current challenges in therapeutic cell production and processing.
The ticket here was to kick the cells into a special “state” that suppressed the expression of some proteins involved in signalling for the building of other proteins. With the MAPK/ERK train derailed, cells remained otherwise normal, but remained pluripotent — capable of being turned into any type of cell — while sustaining a high growth rate. This is early stage research, so don’t expect to see human stem cells by the barrel any time soon, but it does indicate that some of the currently perceived limitations on research and treatment may be overcome with relatively simple changes to the wall stem cells are produced.
Why Pancreatic cancer causes weight loss.
The pancreas is essential, and you have only one. But other than that “you can’t do without it,” issue pancreatic cancer remains one of the most difficult to treat. Intervention, even in cases where tumors are detected early, tends to be massive and hugely impactful (see Whipple Procedure) because it has to be. Not hitting it hard and early, means not getting a second chance.
One of the early symptoms of pancreatic cancer is often overlooked, and sometimes even welcomed — weight loss. That is, until that weight loss leads to weakness and illness. Because weight loss from pancreatic cancer, and other cancers related to the digestive system, isn’t just the body shedding fat, but something that goes under the much less pleasant term of “tissue wasting.” Tissues in the muscles, organs, even bones are gradually broken down for nutrients by a system that can no longer properly extract what it needs from food.
Large loss of weight can occur months before diagnosis of pancreatic cancer. And catching the disease before it causes profound tissue wasting is a significant factor in improving survival. Now researchers from MIT and the Dana-Farber Cancer Institute have come up with a better understanding of the mechanisms behind how pancreatic cancer causes profound weight loss.
Tissue wasting is a multifactorial disease and targeting specific circulating factors to reverse this syndrome has been mostly ineffective in the clinic. Here we show that loss of both adipose and muscle tissue occurs early in the development of pancreatic cancer. Using mouse models of PDAC, we show that tumour growth in the pancreas but not in other sites leads to adipose tissue wasting, suggesting that tumour growth within the pancreatic environment contributes to this wasting phenotype.
The good news here is that early weight loss, which is primarily fat loss, doesn’t seem to affect long-term survivability. Treatment after this initial loss was just as effective. But pancreatic enzymes also regulate the uptake of nutrients like Vitamin D, meaning that a tumor still small and only in the pancreas can trigger systemic issues if it remains unchecked. So … as happy as people may be to shed a few pounds, a significant weight loss that’s no accompanied by a dedicated effort to reduce intake is worth checking out.
Just to give a sample of how inaccessible health researchers can make a sentence when they really really try, I present the opening line from this paper by a team from Switzerland.
Adipocyte development and differentiation have an important role in the aetiology of obesity and its co-morbidities.
Uh huh. Or “Different kinds of fat cells produce different outcomes when it comes to actually making an animal overweight and causing related health problems.” But the interesting thing in this paper is that one of the three kinds of cells in a family that usually come before mature fat cells, actually seems to slow the deposition of fat.
We identify one of these subpopulations as CD142+ adipogenesis-regulatory cells, which can suppress adipocyte formation in vivo and in vitro in a paracrine manner.
Yeah. That’s what I said. The team named these cells “adipogenesis regulators” and confirmed that they work to suppress fat both in lab tests and in animal tests. Mice fed a high-fat diet but given injections of this cell type (weirdly enough, in one half of their bodies) did not become fat (in that half). Human beings also appear to have these cells, although simply identify them required the development of new techniques looking at a single gene within the cells. But whether this will lead to any treatments for obesity is still some time away — so don’t go taking any strange injections. Especially in just one half of your body. Because if there’s anything worse than being overweight, it’s being lop-sided.
IMAGE
This week’s image, as usual comes from Andy Brunning at Compound Interest. Visit his site for a larger, easier to read version along with dozens of other interesting infographics.