Tag Archives: level
Machine learning algorithms are starting to exceed human performance in many narrow and specific domains, such as image recognition and certain types of medical diagnoses. They’re also rapidly improving in more complex domains such as generating eerily human-like text. We increasingly rely on machine learning algorithms to make decisions on a wide range of topics, from what we collectively spend billions of hours watching to who gets the job.
But machine learning algorithms cannot explain the decisions they make.
How can we justify putting these systems in charge of decisions that affect people’s lives if we don’t understand how they’re arriving at those decisions?
This desire to get more than raw numbers from machine learning algorithms has led to a renewed focus on explainable AI: algorithms that can make a decision or take an action, and tell you the reasons behind it.
What Makes You Say That?
In some circumstances, you can see a road to explainable AI already. Take OpenAI’s GTP-2 model, or IBM’s Project Debater. Both of these generate text based on a large corpus of training data, and try to make it as relevant as possible to the prompt that’s given. If these models were also able to provide a quick run-down of the top few sources in that corpus of training data they were drawing information from, it may be easier to understand where the “argument” (or poetic essay about unicorns) was coming from.
This is similar to the approach Google is now looking at for its image classifiers. Many algorithms are more sensitive to textures and the relationship between adjacent pixels in an image, rather than recognizing objects by their outlines as humans do. This leads to strange results: some algorithms can happily identify a totally scrambled image of a polar bear, but not a polar bear silhouette.
Previous attempts to make image classifiers explainable relied on significance mapping. In this method, the algorithm would highlight the areas of the image that contributed the most statistical weight to making the decision. This is usually determined by changing groups of pixels in the image and seeing which contribute to the biggest change in the algorithm’s impression of what the image is. For example, if the algorithm is trying to recognize a stop sign, changing the background is unlikely to be as important as changing the sign.
Google’s new approach changes the way that its algorithm recognizes objects, by examining them at several different resolutions and searching for matches to different “sub-objects” within the main object. You or I might recognize an ambulance from its flashing lights, its tires, and its logo; we might zoom in on the basketball held by an NBA player to deduce their occupation, and so on. By linking the overall categorization of an image to these “concepts,” the algorithm can explain its decision: I categorized this as a cat because of its tail and whiskers.
Even in this experiment, though, the “psychology” of the algorithm in decision-making is counter-intuitive. For example, in the basketball case, the most important factor in making the decision was actually the player’s jerseys rather than the basketball.
Can You Explain What You Don’t Understand?
While it may seem trivial, the conflict here is a fundamental one in approaches to artificial intelligence. Namely, how far can you get with mere statistical associations between huge sets of data, and how much do you need to introduce abstract concepts for real intelligence to arise?
At one end of the spectrum, Good Old-Fashioned AI or GOFAI dreamed up machines that would be entirely based on symbolic logic. The machine would be hard-coded with the concept of a dog, a flower, cars, and so forth, alongside all of the symbolic “rules” which we internalize, allowing us to distinguish between dogs, flowers, and cars. (You can imagine a similar approach to a conversational AI would teach it words and strict grammatical structures from the top down, rather than “learning” languages from statistical associations between letters and words in training data, as GPT-2 broadly does.)
Such a system would be able to explain itself, because it would deal in high-level, human-understandable concepts. The equation is closer to: “ball” + “stitches” + “white” = “baseball”, rather than a set of millions of numbers linking various pathways together. There are elements of GOFAI in Google’s new approach to explaining its image recognition: the new algorithm can recognize objects based on the sub-objects they contain. To do this, it requires at least a rudimentary understanding of what those sub-objects look like, and the rules that link objects to sub-objects, such as “cats have whiskers.”
The issue, of course, is the—maybe impossible—labor-intensive task of defining all these symbolic concepts and every conceivable rule that could possibly link them together by hand. The difficulty of creating systems like this, which could handle the “combinatorial explosion” present in reality, helped to lead to the first AI winter.
Meanwhile, neural networks rely on training themselves on vast sets of data. Without the “labeling” of supervised learning, this process might bear no relation to any concepts a human could understand (and therefore be utterly inexplicable).
Somewhere between these two, hope explainable AI enthusiasts, is a happy medium that can crunch colossal amounts of data, giving us all of the benefits that recent, neural-network AI has bestowed, while showing its working in terms that humans can understand.
Image Credit: Image by Seanbatty from Pixabay Continue reading
What’s the world’s hardest machine learning problem? Autonomous vehicles? Robots that can walk? Cancer detection?
Nope, says Julian Sanchez. It’s agriculture.
Sanchez might be a little biased. He is the director of precision agriculture for John Deere, and is in charge of adding intelligence to traditional farm vehicles. But he does have a little perspective, having spent time working on software for both medical devices and air traffic control systems.
I met with Sanchez and Alexey Rostapshov, head of digital innovation at John Deere Labs, at the organization’s San Francisco offices last month. Labs launched in 2017 to take advantage of the area’s tech expertise, both to apply machine learning to in-house agricultural problems and to work with partners to build technologies that play nicely with Deere’s big green machines. Deere’s neighbors in San Francisco’s tech-heavy South of Market are LinkedIn, Salesforce, and Planet Labs, which puts it in a good position for recruiting.
“We’ve literally had folks knock on the door and say, ‘What are you doing here?’” says Rostapshov, and some return to drop off resumes.
Here’s why Sanchez believes agriculture is such a big challenge for artificial intelligence.
“It’s not just about driving tractors around,” he says, although autonomous driving technologies are part of the mix. (John Deere is doing a lot of work with precision GPS to improve autonomous driving, for example, and allow tractors to plan their own routes around fields.)
But more complex than the driving problem, says Sanchez, are the classification problems.
Corn: A Classic Classification Problem
Photo: Tekla Perry
One key effort, Sanchez says, are AI systems “that allow me to tell whether grain being harvested is good quality or low quality and to make automatic adjustment systems for the harvester.” The company is already selling an early version of this image analysis technology. But the many differences between grain types, and grains grown under different conditions, make this task a tough one for machine learning.
“Take corn,” Sanchez says. “Let’s say we are building a deep learning algorithm to detect this corn. And we take lots of pictures of kernels to give it. Say we pick those kernels in central Illinois. But, one mile over, the farmer planted a slightly different hybrid which has slightly different coloration of yellow. Meanwhile, this other farm harvested three days later in a field five miles away; it’s the same hybrid, but it also looks different.
“It’s an overwhelming classification challenge, and that’s just for corn. But you are not only doing it for corn, you have to add 20 more varieties of grain to the mix; and some, like canola, are almost microscopic.”
Even the ground conditions vary dramatically—far more than road conditions, Sanchez points out.
“Let’s say we are building a deep learning algorithm to detect how much residue is left on the soil after a harvest, including stubble and some chaff. Let’s drive 2,000 acres of fields in the Midwest looking at residue. That’s great, but I guarantee that if you go drive those the next year, it will look significantly different.
“Deep learning is great at interpolating conditions between what it knows; it is not good at extrapolating to situations it hasn’t seen. And in agriculture, you always feel that there is a set of conditions that you haven’t yet classified.”
A Flood of Big Data
The scale of the data is also daunting, Rostapshov points out. “We are one of the largest users of cloud computing services in the world,” he says. “We are gathering 5 to 15 million measurements per second from 130,000 connected machines globally. We have over 150 million acres in our databases, using petabytes and petabytes [of storage]. We process more data than Twitter does.”
Much of this information is so-called dirty data, that is, it doesn’t share the same format or structure, because it’s coming not only from a wide variety of John Deere machines, but also includes data from some 100 other companies that have access to the platform, including weather information, aerial imagery, and soil analyses.
As a result, says Sanchez, Deere has had to make “tremendous investments in back-end data cleanup.”
Deep learning is great at interpolating conditions between what it knows; it is not good at extrapolating to situations it hasn’t seen.”
—Julian Sanchez, John Deere
“We have gotten progressively more skilled at that problem,” he says. “We started simply by cleaning up our own data. You’d think it would be nice and neat, since it’s coming from our own machines, but there is a wide variety of different models and different years. Then we started geospatially tagging the agronomic data—the information about where you are applying herbicides and fertilizer and the like—coming in from our vehicles. When we started bringing in other data, from drones, say, we were already good at cleaning it up.”
John Deere’s Hiring Pitch
Hard problems can be a good thing to have for a company looking to hire machine learning engineers.
“Our opening line to potential recruits,” Sanchez says, “is ‘This stuff matters.’ Then, if we get a chance to talk to them more, we follow up with ‘Not only does this stuff matter, but the problems are really hard and interesting.’ When we explain the variability in farming and how we have to apply all the latest tools to these problems, we get their attention.”
Software engineers “know that feeding a growing population is a massive problem and are excited about the prospect of making a difference,” Rostapshov says.
Only 20 engineers work in the San Francisco labs right now, and that’s on a busy day—some of the researchers spend part of their time at Blue River Technology, a startup based in Sunnyvale that was acquired by Deere in 2017. About half of the researchers are focusing on AI. The Lab is in the process of doubling its office space (no word on staffing plans for that expansion yet).
“We are one of the largest users of cloud computing services in the world.”
—Alexey Rostapshov, John Deere Labs
Company-wide, Deere has thousands of software engineers, with many using AI and machine learning tools in their work, and about the same number of mechanical and electrical engineers, Sanchez reports. “If you look at our hiring 10 years ago,” he says, “it was heavily weighted to mechanical engineers. But if you look at those numbers now, it is by a large majority [engineers working] in the software space. We still need mechanical engineers—we do build green machines—but if you go by our footprint of tech talent, it is pretty safe to call John Deere a software company. And if you follow the key conversations that are happening in the company right now, 95 percent of them are software-related.”
For now, these software engineers are focused on developing technologies that allow farmers to “do more with less,” Sanchez says. Meaning, to get more and better crops from less fuel, less seed, less fertilizer, less pesticide, and fewer workers, and putting together building blocks that, he says, could eventually lead to fully autonomous farm vehicles. The data Deere collects today, for the most part, stays in silos (the virtual kind), with AI algorithms that analyze specific sets of data to provide guidance to individual farmers. At some point, however, with tools to anonymize data and buy-in from farmers, aggregating data could provide some powerful insights.
“We are not asking farmers for that yet,” Sanchez says. “We are not doing aggregation to look for patterns. We are focused on offering technology that allows an individual farmer to use less, on positioning ourselves to be in a neutral spot. We are not about selling you more seed or more fertilizer. So we are building up a good trust level. In the long term, we can have conversations about doing more with deep learning.” Continue reading