by Pete Sonsini
Software may have eaten the world, but artificial intelligence has rapidly climbed to the top of the technology food chain. Its appetite is enormous, and its application will be universal.
These days it’s rare to see a company pitch that doesn’t talk about how they are (or will be) using AI or machine learning, often across many aspects of the business and usually focused on making an ever-growing volume of data more accessible, insightful and actionable. Many of the world’s biggest consumer brands are exploring the space with no end goal other than improving their own products.
Yeah, there’s plenty of hype. But it’s a lot more than hype—it’s the future. AI/ML is without a doubt the next seismic shift in computing, and it will play a greater role in shaping our daily lives than any that came before it. Of course, there is a lot less clarity around how that future will unfold, and investors are placing bets up and down the stack and across verticals.
If I had to pick one killer app for machine learning right now, computer vision would be it. Computers that can see or recognize images will have an impact on every industry you can imagine, but until now it simply hasn’t been possible.
When I first started talking to Reza Zadeh about seeding Matroid a couple years ago, he’d been thinking about starting a computer vision company for nearly a decade. But he had held off, believing that the technology wasn’t yet ready for industry. (He would know—Reza built the machine learning algorithms behind Twitter’s who-to-follow feature (the first product to use machine learning at Twitter), was a technical advisor to Databricks for distributed machine learning, and teaches machine learning algorithms as an Adjunct Professor at Stanford.) But in early 2016, Reza decided the time was right and founded Matroid.
Matroid is a pioneering computer vision platform that allows users to create ‘detectors’ to monitor and identify people and objects in streams of video data. In just two years the company has filed multiple patents, launched a product, raised a Series A round from NEA, and signed on customers in the TV, media, security and semiconductor industries.
I think computer vision will be an important category, likely with multiple big winners (even factoring in Google’s inevitable market share and influence). But I’m also attuned to the risks that accompany great enthusiasm. There is a risk of too much money pouring in, leading to confusion and saturation in the marketplace, and making it hard for winners to break away from the pack.
With computer vision, that risk is amplified by the complexity of the technology and the highly competitive landscape. The stakes are enormous and hundreds of companies (and other entities) are looking for any opportunity to establish a foothold in computer vision. Reza and his team must tread carefully, making certain not to expose too much of the underlying technology to the industry until they have reached critical mass in the marketplace and in new detectors being developed on Matroid’s platform. This reticence can be frustrating to prospective partners and customers (and would-be competitors), but it is necessary in an exploding but early category like computer vision.
While some aspects of Matroid’s technology are not well understood outside of the company (yet), it’s clear that they are building a great deal of momentum in a rapidly growing market. Here are some of the key reasons why I think this category—and this company—will create a lot of value and play a meaningful role in AI maturation:
Timing: The enabling technology has matured or decreased in cost to facilitate computer vision—most critically, compute cycles have reached commodity prices. The market is ready and waiting--- corporations and consumers are eager to incorporate computer vision into their businesses and lives.
Approach: Creating detectors for computer vision isn’t easy, and the demand for these tools far outstrips the capacity of data scientists to deliver them. This is true across the AI landscape: Gartner predicts that by 2020, 85 percent of CIOs will be piloting enterprise AI projects using a combination of buy, build and outsource efforts. Yet by some estimates, as few 10,000 people worldwide have the skills needed to develop the complex algorithms driving AI. That’s why Matroid’s approach, focused on ease-of-use (and re-use) is brilliant—the field of who can create detectors expands exponentially. Further, they have created a marketplace model that allows reuse of detectors built on their platform so that all the world can benefit from what has been created.
Flexible architecture: Today data scientists are using new computing platforms to scale machine learning across many dimensions. Matroid’s detectors are built on TensorFlow, but work with any of the frameworks that are available, such as Torch, Caffe, etc. Their clusters are built with Kubernetes, an extensible open-source platform for managing containerized workloads, so the detectors can run on any cloud, and even on-premises easily. This flexibility is central to Matroid’s mission, because it will enable the marketplace to be that much more robust—in fact, researchers utilizing all types of platforms will converge at Matroid’s Scaled Machine Learning conference at Stanford University later this month, which aims to foster exchange of ideas and methods between the platforms and enable algorithm designers to scale.
Endless applications/use cases: The use cases for Matroid’s highly differentiated technology are practically limitless, from security to marketing to plain old efficiency. With Matroid’s “computer vision as a service,” custom detectors can be created for just about any scenario. A user could look for a type of object or person (or a specific object or person) within a private data set, like a security camera feed or photo collection. An organization could monitor for a vehicle type in a surveillance feed, or a sought-after individual. A detector could monitor live tv for sightings of a favorite celebrity. Or a brand could be alerted when their own products (or competitors’ products) appear in a movie.
Bottom line, it’s an extraordinary technology, led by a world-class team. If Matroid were a movie I could fast forward, I’m pretty sure I’d like the way it ends.