Google's AutoML makes labeled data, not model design, the moat

Google now builds custom image models for anyone who uploads labeled photos. Model design becomes a cloud feature, and labeled data the scarce asset.
Google will now design a custom image recognition model for any business that uploads labeled pictures. Cloud AutoML Vision, introduced on 17 January as a closed alpha, promises a simple model in minutes and a production-ready one in as little as a day, and Disney and Urban Outfitters are among the first companies testing it.
The launch turns model design, the scarcest skill in applied machine learning, into a feature of a cloud account. That moves the margin in computer vision. A startup whose product is a custom classifier for a retailer now competes with a button in the retailer's cloud console, and what stays scarce is the labeled data that button needs and the workflow that keeps producing it.
The technology behind the button explains why few companies could offer it. AutoML Vision rests on transfer learning and on automated architecture design, a method in which one network proposes and tests designs for another. Google's research version of that design loop ran on 500 graphics processors for four days to find image models that reached 82.7% top-1 accuracy on ImageNet, the standard benchmark, and the earlier version of the method had needed 800 processors for 28 days. A startup cannot spread a bill like that over one client. Google spreads it over every customer who uploads a folder of photos. The other large clouds are moving the same way. Microsoft opened a public preview of a custom vision service in May, and Amazon launched SageMaker, a managed service to build, train and deploy models, at the end of November. Cheap model building pays the platforms through storage and computing and hurts any company whose margin came from building models by hand. Training and serving stay on the platform, so a customer that builds there tends to stay there.
For founders the useful question shifts from how good the model is to who owns the pictures and the labels. A retailer's product photos are easy to upload. Medical scans, industrial defect images and satellite tiles are scarce, costly to label and often cannot leave the customer's premises, and that is where a specialist keeps its pricing power, unless the problem can be simulated, as self-play in board games showed last month. Fanaura's interest runs to machine learning companies that own a data pipeline no cloud console can reproduce: labeled examples that grow with every customer, collected inside a workflow the customer already pays for. That is the case worth making when founders bring us a computer vision company.
The objection is that an alpha with no published price proves little, and automated design tuned on everyday photos tends to stumble in narrow, hard domains where hand-built models still win. Google has not said what it will charge. The price at general release and the first side-by-side comparisons on specialist data will show whether AutoML replaces machine learning teams or only the easy half of their work.
Either way the direction is set. The value we look for sits in data a cloud provider cannot ask for, not in an architecture it now designs on request.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building machine learning on data a cloud console cannot reach? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.