Last month Google unveiled Cloud AutoML, a new service, in a bid to achieve its “mission to democratise AI.”
It is explained as artificial intelligence that can design artificial intelligence. The idea is that rather than build your own machine-learning software, Google’s Cloud AutoML will take your training data, figure out what you want and then automatically generate intelligent code for you. The reality is it is an image-classification system.
Google’s chief scientist Fei-Fei Li and their head of R&D Jai Li said “only a handful of businesses in the world have access to the talent and budgets needed to fully appreciate the advancements of ML (machine learning) and AI (artificial intelligence).”
“There’s a very limited number of people that can create advanced machine learning models. And if you’re one of the companies that has access to ML/AI engineers, you still have to manage the time-intensive and complicated process of building your own custom ML model.”
Some have read that as saying, don’t think about doing it yourself, you don’t have any decent machine learning programmers and Google can do the AI coding for you. At a price.
At the moment, Cloud AutoML can only deal with computer vision problems.
Businesses are invited to upload images to train a neural network that is tailored to their interests. For example, a clothes manufacturer may want to feed Google’s cloud pictures of t-shirts, sweaters, skirts, and so on, to create a system that can identify different types of clothing in stock.
The images need to be labelled so the software can learn to recognize objects. Google offers a “human labelling service,” or users can hand label the training data themselves.
The project was started internally in Google’s research wing, dubbed Google Brain. They were looking at ways to automate the process of designing neural network architectures using machine learning. They used a mixture of evolutionary algorithms and reinforcement learning to come up with what’s called neural architecture search, which is basically software automating the task of crafting and testing new machine-learning models.
Now, it appears neural architecture search and transfer learning are being used together to power Cloud AutoML, spitting out customized models on demand. It’s interesting to see exciting research deployed in production, but it’s unclear how much of it is genuine intelligence – and how much is templated code and models that are produced by humans and selected by bots for customers.
No one seems sure if Google creates from scratch a new model for every customer or whether it slightly tweaks a previously trained one for different use cases. It’s also unclear how much this service costs. Is it affordable for smaller businesses? Are customers billed by the complexity of the model needed? Google are yet to say.