Google brings machine learning to online spreadsheets with Simple ML for Sheets

Spreadsheets are widely used by organizations of all sizes for a variety of basic and complex tasks.

While simple calculations and graphs have long been part of the spreadsheet experience, machine learning (ML) has not. ML is often seen as too complex to use, while using spreadsheets is meant to be accessible to every type of user. Google is now trying to change that paradigm for its online spreadsheet program Google Sheets.

Google today announced a beta version of the Simple ML for Sheets add-on. Google Sheets has an extensible architecture that allows users to take advantage of add-ons that extend the standard functionality of the application. In this case, Google Sheets takes advantage of ML technology that Google first developed in the open source TensorFlow project. With Simple ML for Sheets, users do not need to use a specific TensorFlow service, as Google has developed the service to be as easy to access as possible.

“Everything runs entirely in the user’s browser,” Luiz Gustavo Martins, Google AI developer, told VentureBeat. “Your data doesn’t leave Google Sheets and models are saved to your Google Drive for later use.”

Holy sheets, Google’s Simple ML can do what with my spreadsheets?

So what can Simple ML for Sheets do? Two of the beginner tasks in the beta highlighted by Google include the ability to predict missing values ​​or detect abnormal values. Martins said these two beginner tasks are easy for anyone to test the ML waters and explore how ML can benefit their business.

Martins noted that in addition to the beginner tasks, the add-on supports several other common ML tasks such as training and evaluating models, generating predictions, and interpreting the models and their predictions. In addition, since Simple ML can export models to TensorFlow, people with programming experience can use Simple ML models with their existing ML infrastructure.

Overcome the challenges of ML complexity with Simple ML for Sheets

Google Sheets users can take advantage of ML without Simple ML, but it may not be easy for the layman.

“We identified knowledge and lack of guidance as the key factors for non-ML practitioners to easily use ML,” Mathieu Guillame-Bert, a software engineer at Google, told VentureBeat. “Using a classic ML tool, like TensorFlow in Python, is like standing in front of a blank page.”

Guillame-Bert said that using a classic ML tool requires the user to understand programming, ML problem framing, model construction, and model evaluation, among other things. He noted that such knowledge is generally acquired over a long period of time through lessons or self-taught.

In contrast, Guillame-Bert said that Simple ML is like an interactive questionnaire. It guides the user and only assumes basic knowledge about spreadsheets.

Using decision forests to power Simple ML

Martins explained that under the hood, the Simple ML add-on trains models using the Yggdrasil Decision Forests Library. This is the same library that powers TensorFlow decision forests.

“For this reason, once trained in the add-on, the advanced user can export the model to any TensorFlow Serving managed service, such as the TensorFlow Serving on Google Cloud,” said Martins.

Guillame-Bert explained that TensorFlow Decision Forests (TF-DF) is a library of algorithms to train new models. In other words, the user provides samples to TF-DF and receives a model in return. He noted that TF-DF does not come with pre-trained models; However, because TF-DF are integrated into the TensorFlow ecosystems, advanced users can combine decision forests and pre-trained models.

According to published researchthe technology behind TF-DF, which is based on the Random Forests and Gradient-Boosted Trees concepts, works exceptionally well for training models on a tabular dataset, such as a spreadsheet.

Guillame-Bert said Google will work to further improve the add-on’s usability. Google also plans to add new capabilities to Simple ML for Sheets that don’t require user ML knowledge.

“During internal testing, we identified several highly requested tasks that we believe will be popular with users,” said Guillame-Bert. “We hope to get feedback from this public launch to prioritize and design those tasks.”