Research papers arrive too quickly for anyone to read them all, especially in the field of machine learning, which now affects (and produces documents in) virtually every industry and company. This column aims to collect the most relevant recent findings and documents, notably but not limited to artificial intelligence, and explain why they are important.
This week, a startup using UAV drones for forest mapping takes a look at how machine learning can map social media networks and predict Alzheimer's, improve computer vision for space sensors, and other news on the recent technological advances.
Predicting Alzheimer's through language patterns
Machine learning tools are used to aid diagnosis in many ways, as they are sensitive to patterns that humans find difficult to detect. IBM researchers have potentially found such patterns in speech that are predictive of the speaker developing Alzheimer's disease.
The system requires only a couple of minutes of normal speech in a clinical setting. The team used a large dataset (the Framingham Heart Study) dating back to 1948, allowing them to identify speech patterns in people who would later develop Alzheimer's. The accuracy rate is about 71% or 0.74 area under the curve for those of you more statistically informed. It's hardly a certain thing, but current baseline tests are barely better than a coin toss at predicting disease so early.
This is very important because the earlier Alzheimer's can be detected, the better it can be managed. There is no cure, but there are promising treatments and practices that can delay or mitigate the worst symptoms. A quick, non-invasive test of people like this could be a powerful new screening tool and is also, of course, an excellent demonstration of the usefulness of this field of technology.
(Don't read the document expecting to find exact symptoms or anything like that - the range of speech functions isn't quite the kind of thing you can look for in everyday life.)
Cellular networks
Making sure your deep learning network generalizes to data outside of its training environment is a key part of any serious ML research. But few attempt to free a model of data that is completely foreign to it. Maybe they should!
Researchers from Uppsala University in Sweden took a model used to identify groups and connections in social media and applied it (not without modification, of course) to tissue scans. The tissue had been treated so that the resulting images produced thousands of tiny dots representing mRNA.
Normally the different groups of cells, representing the types and areas of the tissue, should be identified and labeled manually. But the graph neural network, created to identify social groups based on similarities as common interests in a virtual space, has been shown to perform a similar task on cells. (See the picture above.)
"We are using the latest artificial intelligence methods - in particular, graphical neural networks, developed to analyze social networks - and adapt them to understand biological patterns and subsequent variations in tissue samples. Cells are comparable to social groupings that can be defined based on the activities they share in their social networks, "said Carolina Wählby of Uppsala.
It is an interesting illustration not only of the flexibility of neural networks, but of how structures and architectures repeat themselves at all scales and in all contexts. As outside, so inside, If you want.
Drones in nature
The vast forests of our national parks and timber farms have countless trees, but you can't put "countless" on records. Someone has to make an actual estimate of the growth of the various regions, the density and types of trees, the variety of diseases or fires, and so on. This process is only partially automated, as aerial photography and scans only reveal so much, while ground observation is detailed but extremely slow and limited.
Treeswift aims to take a middle ground by equipping drones with the sensors they need to both navigate and accurately measure the forest. By flying through much faster than a person walking, they can count trees, observe problems, and generally collect a ton of useful data. The company is still in its infancy, having exited the University of Pennsylvania and acquired an SBIR grant from the NSF.
"Companies are increasingly looking for forest resources to combat climate change, but you don't have an offer of people who are growing to meet that need," Steven Chen, co-founder and CEO of Treeswift and PhD student in Computer and Information Science (CIS) at Penn Engineering said in Penn news. "I want to help every forester do what they do more efficiently. These robots will not replace human jobs. Instead, they are providing new tools to people who have the intuition and passion to manage our forests. "
Another area where drones are making a lot of interesting moves is underwater. Autonomous deep-sea submarines are helping to map the seafloor, monitor ice shelves and track whales. But they all have a bit of an Achilles heel as their data needs to be periodically collected, uploaded and retrieved.
Purdue engineering professor Nina Mahmoudian has created a docking system by which submarines can easily and automatically connect for power and data exchange.
The boat needs a special nosepiece, which can find and connect to a station that establishes a secure connection. The station can itself be an autonomous vessel or a permanent function somewhere - what matters is that the smaller vessel can pit to reload and debrief before continuing on. If it is lost (a real danger at sea), your data will not be lost with it.
You can see the configuration in action below:
https://youtu.be/ kS0-qc_r0
I play in theory
Drones may soon become fixed elements of city life, although we are probably a long way from the automated private helicopters that some seem to think are just around the corner. But living under a highway with drones means constant noise, so people are always looking for ways to reduce turbulence and the resulting sound from wings and propellers.
Researchers from King Abdullah University of Science and Technology have found a new and more efficient way to simulate airflow in these situations; fluid dynamics is essentially as complex as you make it, so the trick is to apply your computing power to the right parts of the problem. They were only able to reproduce the flow near the surface of the theoretical aircraft in high resolution, finding beyond a certain distance it made no sense to know exactly what was going on. Improvements to reality models don't always have to be better in every way - after all, results are what matter.
Machine learning in space
Machine vision algorithms have come a long way and, as their efficiency improves, they begin to be implemented at the edge rather than in data centers. In fact, it has become quite common for camera-based objects like phones and IoT devices to do some local ML work on the image. But in space it's another story.
Running machine learning work in space was until recently simply too costly from an energy standpoint to consider. This is the power that could be used to capture another image, transmit data to the surface, etc. HyperScout 2 is exploring the possibility of working with machine learning in space, and its satellite began immediately applying computer vision techniques to the images it collects before sending them. down. ("Here is a cloud - here is Portugal - here is a volcano ...")
There is little practical benefit for now, but object detection can easily be combined with other features to create new use cases, from saving energy when no objects of interest are present, to passing metadata to other tools that might work. better if informed.
Inside the old, outside the new
Machine learning models are great for making educated guesses, and in disciplines where there is a large backlog of unsorted or poorly documented data, it can be very helpful to let an AI take the first step so graduate students can use their time more productively. The Library of Congress is doing this with old newspapers, and now Carnegie Mellon University libraries are getting into the spirit.
The CMU photo archive of millions of articles is being digitized, but to make it useful for historians and curious browsers it needs to be organized and tagged, so computer vision algorithms are put to work by grouping similar images, identifying objects and places , and perform other valuable basic cataloging activities.
"Even a partially successful project would greatly improve the collection's metadata and could provide a possible solution for metadata generation if the archives were ever funded to digitize the entire collection," said CMU's Matt Lincoln.
A very different project, but one that seems somehow connected, is this work of a student of the Escola Politécnica da Universidade de Pernambuco in Brazil, who had the brilliant idea of trying to touch up some old maps with machine learning.
The tool they used takes old dashed maps and attempts to create a kind of satellite image based on them using a Generative Adversarial Network; GANs essentially attempt to induce themselves to create content that they cannot distinguish from the real thing.
Well, the results aren't what you might call entirely convincing, but they are promising nonetheless. Such maps are rarely accurate, but that doesn't mean they're completely abstract - recreating them in the context of modern mapping techniques is a fun idea that could help these places seem less distant.