AI United States

Google Cloud’s Jia Li on AutoML’s First 30 Days

Who is using AutoML Vision, the “simple, secure and flexible ML service that lets you train custom vision models for your own use cases” that Google launched last month?

Who is using AutoML Vision, the “simple, secure and flexible ML service that lets you train custom vision models for your own use cases” that Google launched last month? We asked Jia Li, Head of R&D of Cloud AI and Senior Director at Google.

Li says AutoML now has some 10,000 registered users, ranging from startups to big corporations. She divides potential clients into three types, “the first is AI-savvy companies with loads of data, they need infrastructure such as Tensorflow to train their own models; second is companies with limited expertise and data, they can choose API without training their own models; third is companies with limited expertise, but with ideas and data, that want to build their own models. AutoML can help them customize their models by simply inputting their images without marking training data, designing algorithms, or tuning parameters.”

AutoML users can drag-and-drop to upload images, no AI technical expertise is required. AutoML does this with a combination of three core technologies — neural architecture search technology, learning2learn, and transfer learning — which automate the process of selecting the right networks to use, finding hyperparameters for best performance, and applying the model to different use cases directly from Google Cloud.

download (1).gif

Screen Shot 2018-02-07 at 1.44.08 PM.png

Li explains: “Transfer learning is easy because it generates models in seconds. Learning2learn has a higher cost with no fixed architecture, and it takes up to a day to generate models. But even this is shorter than traditional ways of training.”

Google reports on their blog that “early results using Cloud AutoML Vision to classify popular public datasets like ImageNet and CIFAR have shown more accurate results with fewer misclassifications than generic ML APIs.” AutoML’s codes run better than those written by engineers; while for labeling objects in an image AutoML achieves 42% accuracy compared to man-made models’ 39%.

Li identifies retail and medical imaging as major use cases. In the month since AutoML’s launch Google has consulted with many potential clients in the clothing industry. Clothing with the same color or pattern will have for example different necklines, cuffs etc. Retailers can use AutoML to define their own custom product classifications for such features.

The debut version of AutoML only supports computer vision models, but features such as speech, translation, video, and NLP are coming soon.

Fei-fei Li and Jia Li have said that Google’s mission is to “democratize AI” by providing companies who can’t afford their own AI talent a chance to hop on the fast track.

While AutoML’s emergence is good news for businesses, one fear is that the tech giants’ cheap and generalized solutions will raise the bar for startups and reduce market penetration opportunities for their solutions.

The industry has so far responded favourably to AutoML. AI engineers are mostly happy that it eliminates laborious parameter tuning procedures. Businesses meanwhile can now introduce AI into their operations without the high cost of hiring AI engineers and data scientists. Google’s AutoML boasts a growing client list that includes Urban Outfitters, Disney, and the Zoological Society of London and it is likely that Google and Microsoft will continue to expand in this space.


Journalist: Meghan Han | Editor:  Michael Sarazen

0 comments on “Google Cloud’s Jia Li on AutoML’s First 30 Days

Leave a Reply

Your email address will not be published.

%d bloggers like this: