A lot of companies are trying to make it easier to use artificial intelligence, but few are making it as simple as Lobe. The startup, which launched earlier this year, offers users a clean drag-and-drop interface for building deep learning algorithms from scratch. It’s mainly focused on machine vision. That means if you want to build a tool that recognizes different houseplants or can count the number of birds in a tree, you can do it all in Lobe without typing a single line of code.
Company co-founder Mike Matas told The Verge that Lobe isn’t designed to compete with software used by machine learning professionals (tools like PyTorch and TensorFlow). Instead, it’s built to give amateurs an easy way in. “People have ideas they want to try in machine learning but don’t have the right way to prototype them,” says Matas. Lobe lets them take that first step without any training, making deep learning more accessible to professionals in a diverse range of fields, from architecture to astronomy. “We want to have a broad appeal,” he adds.
Lobe and tools like it offer access to deep learning without coding
To put these claims to the test, I tried out Lobe for myself. I built a deep learning algorithm that recognizes scrabble pieces through a webcam. Show it a letter “a” and it says “a,” show it a “b” and it says “b,” and so on. (It’s groundbreaking stuff.) To do so, all I had to do was go through steps common to most machine learning applications, starting by collecting a dataset (pictures of scrabble pieces), labeling that data (sorting them by letter into different folders), then letting a neural network sift through the pictures and slowly learn the shapes that make up each letter.
For this sort of task, Lobe couldn’t be easier to use. You don’t need to download anything; just load up Lobe.ai in your browser, sign in with a Google account, and you’re ready to go. You select a template (in this case, matching images with predefined labels), drop in your data from your desktop, and let it crunch through the information for you. There are only a few templates at the moment, but Lobe’s creators say they plan to expand this by adding new neural network architectures over time and creating a community where users can share their best models.
The different parts of the neural network appear on-screen as boxes (called “lobes,” naturally), which are joined together by lines like a flowchart. You can also look “inside” each of the nodes and make adjustments to how they process data. When you’re finished tinkering, you can export the final model to a number of different platforms, including Google’s TensorFlow and Apple’s iOS-based CoreML.
Lobe.ai is in beta right now, so there were a few rough patches in the design and a lot of unintuitive detail. But the overall look and feel is impressively slick, thanks in part to Matas’ background. A designer by trade, he previously worked at Apple doing UI work for the first iPhone and first iPad. Then, he moved on to Nest to work on the company’s thermostat design, and then Facebook, where he and his team built the gorgeous (but doomed) Paper app.
When building our scrabble detector, I did have some teething difficulties (not feeding the system enough data was the main one), but it was nothing a bit of idle tinkering couldn’t solve. Before long, I had a little scrabble-piece-recognizer program running in my browser.
Obviously, what I built was incredibly simple. But to understand the potential of this sort of tech, don’t think of it as AI; think of it as an obedient monkey that’s house-trained and you’ve taught to recognize a visual cue of your choice. Our monkey friend can’t do much but go “eee eee eee” whenever it spots whatever it is you asked it to look out for, but that’s still a handy monkey to have around. With the right training materials and a little patience, you can teach it to do all sorts of useful things, like recognizing the difference between malignant and benign moles or watching the baby monitor to see if your infant daughter has escaped her crib. You can even get your monkey to recognize different sounds if you convert the audio signals into visual data as a sound wave. And, yes, these are all real applications built using Lobe.
Of course, as much as Lobe’s simplicity is appealing to machine learning amateurs, that doesn’t mean it’s without flaws. Some experts argue that tools like this flatten out their discipline to an unhelpful degree, simply replacing the coding process with a visual stand-in that doesn’t actually teach them how to build quality algorithms. Jeremy Howard is a data scientist and entrepreneur who co-founded Fast.AI, a research lab that makes deep learning more accessible through tools and tutorials. He says he’s seen it all before.
“For some reason, every year or two, someone comes along and designs a machine learning training system that involves dragging boxes from a toolbox and drawing them together with lines. I haven’t quite figured out why this is so appealing to build, but it is,” says Howard. “They always get a certain amount of column inches because people outside the community think they’ve made machine learning easy, but it doesn’t.”
What if Lobe oversimplifies machine learning?
Howard says that there isn’t really much point to these visual interfaces; the process is essentially the same as writing code, but it’s “more awkward, takes longer, and you see less on the screen at once.” He points out that to build anything other than the most basic application (like I did), you still have to know which components you want to use, how to connect them together, and so on. But instead of making that information as accessible as bare code, it’s wrapped up lines and boxes. “The hard bit of data science is not the typing,” says Howard. “It’s knowing what to write.”
He also suggests that some of Lobe’s more complex settings are there for the aesthetics, rather than actual functionality. “They show you changing the architecture of the model … but no one hand-writes architecture. Nobody,” says Howard. “The idea of somehow doing that by typing numbers into boxes shows a complete lack of understanding of what people need to do.”
If you agree with Howard’s criticisms, then Lobe may seem like an overcomplicated way to get people started in machine learning. (Compared to free alternatives, it’ll also cost you, though Matas says pricing has not yet been finalized.) But even if visual tools like this are just skins for existing software, there’s still benefit to them. The pop culture image of code — those scrolling, never-ending lines of impenetrable hieroglyphs — is an intimidating one. Lobe makes getting started less scary, and, in turn, it could help induct professionals who would benefit from the latest machine learning techniques.
Jean-Olivier Irisson is a good example of this phenomenon. An associate professor at France’s Laboratory of Oceanography in Villefranche-sur-Mer, Irisson tells The Verge that he’s not a total stranger to coding, but he’s had no direct experience with deep learning before using Lobe. Now, he’s part of the closed beta and is using the company’s software to help classify images of plankton.
Irisson says he and his colleagues are all aware of new deep learning techniques that outperform traditional approaches in their field, “but deep learning is progressing at a pace that we, as non specialists, cannot follow.” He says Lobe means he can get started with the latest neural network architecture without having to buy new hardware or come to grips with a new coding framework. “It allows me to focus on reading papers, understanding the concepts, and applying those I find promising,” he says.
Like myself, Irisson says he found certain parts of Lobe’s UI glitchy, but again, this is beta software. Overall, he says, the experience was intuitive. “I think a tool like Lobe can help non-specialists open the ‘black boxes’ these [deep learning] models are thought to be and understand them; it certainly helped me,” he told The Verge over email.
This seems to chime with Matas’ vision for Lobe: a tool that lets people start building AI applications as speedily as possible. “If you’ve got a concept for something, and you have the data, you can quickly train a machine learning model on it,” says Matas. “And if it’s possible and works well with the 1,000 examples you have, you can find where to improve and you have the motivation to keep going.”
If AI really is going to change the world, then it seems obvious that the more people who get involved, the better — especially people outside the tech world. There are professionals in scientific fields who may not feel they have the time to learn to code, but they can play around in their browser to get a feel for things. With that in mind, Lobe looks like a perfect recruitment tool for a machine learning revolution.