Google Tulip: The Technical Details

How we use machine learning to improve the well-being of flowers

with Lee Boonstra, Full Stack Plant Engineer, Google Tulip

It is our pleasure to announce Google Tulip, our machine learning technology to improve the profitability of Dutch agriculture and the well-being of Dutch cash crops.

Introducing Google Tulip

How we did it

The Netherlands produce nearly 12.5 billion flowers per year. Dutch tulips are among the tallest flowers in the world, and build strong communities within the fields. Researchers have shown that the tulips communicate with each other through their root systems and are able to share resources. Famously, a little red tulip was able to plug a leak in the seawall that protects Holland, thanks to the early warning system perfected by the network of roots that stretches all across the Netherlands.

Many people objectify tulips and want to just display them in vases in a way that is disconnected from the roots that supply so much meaning to their lives. At Google, we believe in organizing the world’s information and the information inherent in the root network was appealing to us. Thanks to collaboration with Wageningen University & Research and libraries of past audio files, we have digitized flower communications over the centuries and have built a machine learning system to identify what the tulips are communicating. By doing so, we have been able to improve the lives of the tulips (at least up to the point they are cut and shipped).

System architecture

The training architecture was quite simple. We were able to use Google Cloud Speech to Text and Auto ML Natural Language to train the machine learning models without having to write any code.

Training architecture for Google Tulip takes advantage of some prebuilt models and Auto ML Natural Language which employs neural machine translation and neural architecture search.

Carrying out real-time predictions was a bit more challenging, especially because of connectivity problems tying more than a million tulips together. We use Cloud IoT Core to collect the audio data from individual tulips, and carry out predictions on Kubeflow Pipelines “on-premises”. The requests from the tulips are then acted upon by human overseers, as shown in this video..

Real-time serving architecture for the predictions

Deployment

When the flowers need more sun, we simply call out to the Google Wind API to make that happen. This is the technology that we announced two years ago on this exact day to employ machine learning to predict wind and ensure clear skies in the Netherlands, and the API is quite easy to integrate.

Besides the benefits to Dutch agriculture, an unanticipated side-effect of Google Tulip has been to add close captioning of the flowers’ Tulipish to classic Bollywood dance sequences filmed in the Netherlands. Watch for lip-syncing tulips in Bollywood movies this year!

Tulip Translator on Google Home

Talk to your Tulip Translator on Google Home or Google Assistant. Available April 1.

And on April 1st, 2019, only: look for it on your Google Home device, simply by saying, “Hey Google, Talk to Tulip Translator”. Ask the flower what it needs to grow faster.

Do It Yourself Version

If you would like to try the technology in your garden on your own flowers, you can use our Do-It-Yourself kit that employs Google Cloud tools. Head over to our GitHub repo to check out sample code and tutorials to build a DIY version:

In our GitHub repo, you will find sample code to build a do-it-yourself version that starts from your phone camera.

The DIY version is best for users growing less than 1 million flowers: even a single flower will do. For convenience, the DIY version uses photo recognition instead of root systems, and takes advantage of Cloud Vision AutoML to understand the flowers.

Next steps:

  • Watch the video.
  • Say to your Google Home or Google Assistant (e.g. your Pixel phone): “Hey Google, Talk to my Tulip Translator”
  • Check out the DIY code on GitHub.
  • Share this article!