Google AutoML – Two Real Life Examples of Google’s Automated Machine Learning Tool in Action
Overview
- Google’s AutoML is seeing widespread adoption in the data science world
- In one case, a data scientist build a model to predict the restaurant where a bowl of noodles was made, with 94.5% accuracy
- Another instance has seen a popular company increase their model accuracy by over 16% from their existing model
Introduction
Google’s AutoML has been improving steadily since it’s release as more and more data scientists are incorporating it to solve real life problems. This week, Google published a couple of examples on their blog detailing the use cases. Let’s look at them below.
The Restaurant Location Recognizing Model
If someone asked you to eat a bowl of noodles from a restaurant you haven’t visited before, how likely are you to identify which restaurant it comes from? Sounds like an impossible question, right?
Not for Google’s AutoML tool. A machine learning model, trained using AutoML, was able to identify the restaurant with 95% accuracy by looking at the image of the noodle bowl. The model can take apart and analyse minute details of the image and is able to predict which restaurant it was made in. Incredible!
The developer of the model, Kenji Doi, collected 48,244 photos of noodle bowls from different Ramen Jiro locations (Ramen Jiro is a popular restaurant franchise in Japan). Then, Kenji removed photos that was unsuitable for training, such as duplicates. He then labelled 48,000 photos (1170 photos from 41 Ramen Jiro locations).
Kenji then uploaded his dataset into AutoML and found that the trained model gave him a 94.5% accuracy (94.8% precision and 94.5% recall). In simple words, the model was able to recognise the location where the bowl of noodles was prepared an incredible 94.5 times out of 100!
You can check out the confusion matrix below for the model. Note that the rows represent the actual shop and the columns are the predicted shop.
Classifying Branded Goods
In another instance, Mercari leveraged AutoML to classify images with its brand name.
The company has recently launched a new application to sell branded goods in Japan. According to Mercari, they have been “developing their own ML model that suggests a brand name from 12 major brands in the photo uploading user interface”. Their current model uses transfer learning on TensorFlow and gives an accuracy of around 75%.
But then data scientists at Mercari decided to try their hand at AutoML. The results blew their old model out of the water. They trained the model on 50,000 images and achieved an accuracy of 91.3%. An incredible 16% increase on their already existing model!
Check out the confusion matrix for Mercari’s AutoML model below:
Our take on this
So far, AutoML had been labelled as one of the best tools out there for people who did not have the technical data science skills. But as you can see in these examples, data scientists are also leveraging it in real-life cases to build and improve their models.
AutoML takes out the need for hyperparameter tuning and data augmentation. All you need to do is label images, upload them and AutoML takes care of the rest. You can read more about the tool in our post here.
Have you used AutoML so far? Tell us about your experience in the comments below.
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!
Senior Editor at Analytics Vidhya. Data visualization practitioner who loves reading and delving deeper into the data science and machine learning arts. Always looking for new ways to improve processes using ML and AI.
