When he was 13, Rajeev Dutt, inspired by science-fiction authors such as Isaac Asimov, envisioned a world with ubiquitous artificial intelligence. His startup, DimensionalMechanics, is one of many companies large and small trying to deliver on that vision.
The company has built a set of software tools meant to make it easier for non-experts to build systems that can approximate human capabilities, such as identifying objects in photos, filtering adult content, or writing headlines.
DimensionalMechanics is trying to distinguish itself by providing a “horizontal platform” on which companies can build machine learning models, while many other A.I. and machine learning startups have focused on solutions tailored to specific vertical industries. (That said, DimensionalMechanics is favoring media and entertainment out of the gate through a partnership with GrayMeta, which provides metadata, tagging, content management, and search services to movie studios, broadcasters, and law enforcement.)
The idea of democratizing access to artificial intelligence—more specifically the machine learning algorithms that can do things such as classify images or predict the performance of headlines—has been bandied about more frequently in recent years. It’s a pillar of the A.I. strategies of many of the major players, such as Microsoft (NASDAQ: [[ticker:MSFT]]), which has released a steady drumbeat of tools and training datasets, and smaller companies, such as Turi, which was acquired by Apple (NASDAQ: [[ticker:AAPL]]) last year. DataRobot, a leading Boston-area startup that raised $54 million earlier this year, has a similar mission, as does Nutonian, the company DataRobot it just acquired.
These tools are meant to appeal to businesses that have amassed troves of data and now want to do something useful with it. But smaller companies struggle to attract people with the requisite skills to build machine learning systems in-house, and Dutt argues they shouldn’t have to: Just as a generation ago, when companies transitioned away from building their own bespoke database systems, opting instead for off-the-shelf technology, there’s no good reason a widget builder today should be building their own machine learning system, he says.
Add to that the feverish A.I. and machine learning hype cycle mounting over the last 18 months and you’ve got a potential business opportunity for companies like DimensionalMechanics, based in Bellevue, WA, and founded in 2015. The 11-person company last year raised $6.7 million in debt and equity, all from angel investors, led by Kent Johnson—a surprisingly large sum without going to traditional venture capital firms. Dutt says another funding round is planned this fall, and he will be seeking support from VC firms.
Whether the “horizontal” approach in machine learning will take off is an open question. Shivon Zilis, partner at early stage venture firm Bloomberg Beta, described startups such as Turi as “Peter Pans” in a blog post on the machine intelligence landscape earlier this year.
“Established companies struggle to understand machine intelligence technology, so it’s painful to sell to them, and the market for buyers who can use this technology in a self-service way is small,” Zilis wrote. “Then, if you do understand how this technology can supercharge your organization, you realize it’s so valuable that you want to hoard it. Businesses are saying to machine intelligence companies, ‘forget you selling this technology to others, I’m going to buy the whole thing.’
“This absence of a market today makes it difficult for a machine intelligence startup, especially horizontal technology providers, to ‘grow up’—hence the Peter Pans.”
Dutt says DimensionalMechanics has something unique built into its new software, NeoPulse A.I. Studio, released Thursday: “A machine that’s building other machines, if you will.” Dutt says the software is “a way to allow non-specialists to build up A.I. models in a fraction of the time, and also allow A.I. specialists to spin up solutions much faster than they can.”
To demonstrate, Dutt says DimensionalMechanics created a photo ranking tool that predicted how well a given image would perform on Instagram. He says the algorithm performed as well as a human producer in chosing images to illustrate a story. Many companies make similar claims, but Dutt says the NeoPulse tools allows for models such as this to be built in weeks or less, instead of months.
Part of the toolset includes a programming language DimensionalMechanics built for this purpose, NeoPulse Modeling Language, Dutt says. “We found that no other language out there could really help us accelerate the process,” he says. “None of them were really intuitively designed for deep learning.”
Associated with the language is a compiler that automates the process of designing and selecting “the optimal deep learning architecture” for a given task. Dutt calls this “the oracle … an A.I. behind the A.I.”
“You don’t need to know anything,” Dutt says. “There is a keyword, ‘auto,’ which you use in your code and that just instructs the A.I., which we call the oracle, behind A.I. Studio, to say, ‘Hey, please build me a deep learning architecture.’”
The result is a neural network tailored to solve the specified problem, built without experts in a fraction of the time, he says.
But as with many, if not most, A.I. systems, there’s not always a clear explanation for the answers that NeoPulse produces. It’s known in the field as the black box problem. You input a data set and get a solution, but “understanding why is a little bit harder,” Dutt says.
For certain problems, the company’s systems allow users to inspect certain layers of a deep learning model to see if, say, an image classifier is keying off of shape, color, orientation, and the like. “There are many other problems where we won’t have a clue,” he says, adding that the company is working on it—as is the machine learning field at large.
One potential solution involves letting users prioritize the various factors a model weighs in coming to its answers. This is typically an automated process in A.I. Studio, he says.
The black box problem raises broader questions for DimensionalMechanics and others looking to democratize A.I.: If you know nothing about A.I. systems, and you’re making business decisions based on a largely inscrutable algorithm—inscrutable to both the experts building it and the laity to whom it’s being sold—how can you trust it? And does that even matter?
Dutt says that in some industries, it does. “We have spoken to a couple of investment banks and they’re interested broadly in the technology, but they’re wary of machine learning in general because a, it’s probabilistic, and b, you’re not always able to understand the reason why a decision was made,” he says.
That said, he believes we’ve become accustomed to that uncertainty, as long as it works. Dutt, who named his company and his dog (Feynman) in homage to his background in and love for theoretical physics, recalls his initial shock when a professor told him during his undergraduate studies that there was no good explanation for certain phenomena—the behavior of magnetic fields on the surface of the sun, in this instance—beyond what the equations tell us. (Perhaps NASA’s Parker Solar Probe, launching in 2018, will provide some answers to that particular question.)
“This is our general attitude toward technology as human beings,” Dutt says. “We tend to accept things if they solve our problem. If they don’t, then we don’t use it. And very rarely do we ask the question, ‘Why?’”
He does note, however, that the data fed into a model matters greatly. “If your data is biased, your model will be biased as a result of that,” Dutt says. He adds that biased data is a persistent problem, and one that can be helped by easing the process of building machine learning models, allowing companies to more quickly test datasets and correct bad data.
Dutt says 22 customers have pre-ordered A.I. Studio. The company’s business model involves transactions. Say someone builds a great facial recognition model on the NeoPulse platform, which includes an app store. An app in that store might use the facial recognition model, paying the creator of the model each time it calls on the model’s API. Each time it does, DimensionalMechanics takes a cut of the resulting transaction.
The company also offers an on-premises version of its software on a freemium model, giving customers free license to build models and then charging them as they are used.
Photo credit: Spiral staircase photo by Ludde Lorentz, accessed via Unsplash, used under the site’s nonexclusive copyright license.
