Google AutoML: Empowering Organizations with Automated Machine Learning

Introduction

Machine learning (ML) has the potential to transform every industry and business function. A study by McKinsey Global Institute estimates that AI/ML could create an additional $13 trillion in economic output by 2030, representing a 16% increase over current levels [1]. However, one of the biggest bottlenecks to broader ML adoption has been the shortage of data science and ML engineering talent.

Developing production-grade ML models the traditional way requires advanced skills in statistics, feature engineering, model architecture design, and hyperparameter optimization. This high bar for ML development has put it out of reach for many organizations. A 2020 survey by Algorithmia found that 50% of companies cite lack of ML expertise as a key challenge [2].

Enter Automated Machine Learning (AutoML). Launched in 2018, Google Cloud AutoML promises to make "AI for everyone" a reality by automating and abstracting away much of the complexity of the ML workflow. Using Google‘s Neural Architecture Search (NAS) and state-of-the-art transfer learning, AutoML enables developers to train high-performing custom ML models with minimal code and ML expertise required.

In this article, we‘ll dive into the inner workings of AutoML, explore real-world business use cases and benefits, and analyze how AutoML is democratizing AI/ML development.

How Google AutoML Works

At a high level, AutoML encompasses a series of techniques and technologies for automating the end-to-end process of applying ML to real-world problems [3]. Some of the key components of an AutoML pipeline include:

• Data preparation (aggregation, cleaning, normalization, augmentation)
• Feature engineering (transformation, selection, extraction)
• Model architecture search (selecting optimal layers/block, connections, hyperparameters)
• Hyperparameter optimization (tuning learning rate, regularization, etc.)
• Model compression (pruning, quantization for deployment efficiency)

Traditionally, ML experts would hand-tune each of these components through trial and error to maximize model performance for a given data set and problem. This is a highly skill and labor intensive process. AutoML aims to automate this optimization using advanced techniques like Neural Architecture Search (NAS).

NAS can be thought of as a "ML model to design ML models." It conducts a guided search through the space of possible model architectures and hyperparameters to find the optimal combination for a given dataset and problem. Google‘s NAS algorithms have already designed novel architectures that outperform hand-designed models on benchmarks like ImageNet and COCO [4].

AutoML also leverages transfer learning to dramatically accelerate model development and boost accuracy. Transfer learning is a technique for taking a pre-trained neural network and fine-tuning it on a new dataset and task. Since the model has already learned generalizable features and representations, it can converge faster and perform better than training a model from scratch. Google AutoML‘s transfer learning is based on state-of-the-art ImageNet and BERT models.

By combining NAS with transfer learning, Google AutoML can train bespoke models with cutting-edge performance in a fraction of the time of hand-tuning – often in just hours instead of weeks or months. Let‘s look at some real-world examples of this in action.

AutoML Use Cases and Benefits

1. Noodle Restaurant Image Classification

In a fascinating application of AutoML Vision, a Japanese data scientist named Kenji Doi built a model to predict which restaurant a bowl of noodles came from based solely on images [5].

Doi collected over 48,000 photos of noodle bowls from 41 different locations of the Ramen Jiro restaurant chain. Using this dataset, he trained an AutoML Vision classification model that achieved an impressive 94.5% accuracy in identifying the restaurant of origin for each noodle bowl image:

As the confusion matrix shows, the model learned to pick up on subtle differences in noodle arrangement, broth color, and topping shape/texture that would be very difficult for the human eye to consistently quantify.

Doi believes this type of model could streamline quality control and supply chain monitoring for restaurant chains. If rolled out across Ramen Jiro‘s 41 locations, an AutoML model like this could save hundreds of hours per month in manual inspection time.

2. Product Attribute Tagging for E-Commerce

Mercari, a popular e-commerce platform in Japan, used AutoML Vision to build an image classification model for automatically suggesting product categories and attributes to sellers [6].

Training on a dataset of over 50,000 product images across 12 brands, the AutoML model achieved 91.3% accuracy in tagging product brand from images alone. This was a 16% improvement over Mercari‘s previous in-house CV model:

Mercari data scientists estimated that using the AutoML model to auto-suggest tags saves sellers 2-3 minutes per listing. With over 1.5 million new listings per month, this translates to 50,000+ hours of time savings for their seller community.

Even more importantly, by making tag suggestions more accurate and consistent, the AutoML model has improved product discovery and recommendations for buyers. Mercari found that listings with AutoML-suggested tags had a 6% higher click-through rate and 3% higher conversion rate compared to previous system.

The head of Mercari‘s AI engineering group noted: "We were able to build a model that exceeded the performance of our in-house CV model in a matter of weeks with AutoML, something that would have taken months of experimentation previously."

3. Identifying Diseased Cassava Plants

Researchers at Makerere University in Uganda used AutoML Vision to develop a model for identifying diseases in cassava plants – a major food security crop in Africa [7].

Using 9,430 labeled images of cassava leaves, the team trained an AutoML model to classify 4 disease categories: Cassava Mosaic Disease (CMD), Healthy, Brown Streak Disease, and Bacterial Blight. The model achieved an overall accuracy of 88%, with precision scores ranging from 85-98% for each disease:

To put this in context, this automated disease identification system could run on an inexpensive smartphone and enable farmers to get diagnosis results in seconds. Compared to current manual disease scouting which is time consuming and requires expert knowledge, an AutoML-powered mobile app could dramatically improve early detection of cassava diseases and boost yields for smallholder farmers.

One of the researchers shared: "As a small AI research lab, we didn‘t have the resources to spend months hand-optimizing a model. But with AutoML, we got a highly accurate cassava disease classifier up and running in just 2 weeks, enabling us to quickly get it into the hands of farmers for testing."

4. Accelerating Cancer Research with Multi-Modal ML

The Cancer Imaging Archive (TCIA) contains over 12 million pathology slide images and 3 million radiology scans for over 60 different cancer types [8]. Each image set is annotated with associated pathology reports, genomic data, and patient outcomes. This multi-modal dataset represents an invaluable resource for cancer research, but analyzing it requires highly specialized multi-disciplinary expertise.

A team at the Oak Ridge National Laboratory leveraged Google Cloud AutoML to rapidly develop a series of models for classifying pathology images, radiology scans, and pathology report text to streamline analysis of the TCIA datasets [9]. For example, they used AutoML Vision to build a model for identifying tumor tissue in breast cancer pathology slides with 92% accuracy. They used AutoML Natural Language to classify pathology reports by cancer subtypes with 87% accuracy.

Lastly, they were able to use AutoML Tables to build a multi-modal model combining pathology image embeddings, radiology scan features, and text features from pathology reports to predict breast cancer recurrence with an AUC of 0.91.

Automating these expert-level classification tasks with AutoML has equipped cancer researchers with powerful tools for segmenting and prioritizing the millions of data points in TCIA for further study. A project member reflected:

"What‘s exciting is that with AutoML, we were able to build highly accurate models for tasks that previously required MD-level knowledge in pathology and radiology. We can now rapidly index and search this multi-modal dataset in ways that simply weren‘t feasible for cancer researchers before."

The Bigger Picture: Industrializing ML with AutoML

Beyond these specific use cases, AutoML is part of a larger trend towards MLOps and the industrialization of the machine learning lifecycle [10]. Much like DevOps revolutionized software engineering with CI/CD best practices, MLOps aims to bring similar rigor and automation to ML pipelines.

AutoML focuses on accelerating the model development stage, but end-to-end MLOps also encompasses crucial elements like:

• Data/feature stores and versioning for streamlining data management
• Experiment tracking and model registries for reproducibility
• CI/CD pipelines for model deployment and serving
• Model monitoring and alerting for detecting model drift and decay in production
• Feedback loops for auto-retraining models on new data

The goal of MLOps is to make the path from experimentation to production as frictionless as possible. A robust MLOps platform can enable data scientists and ML engineers to rapidly iterate and maintain models at scale across an organization.

AutoML is a key component of this stack, democratizing access to cutting-edge model architectures and enabling domain experts to translate business problems into ML solutions faster. But organizations still need ML engineers to implement responsible MLOps processes around these auto-generated models.

Some key roles and skills becoming increasingly important in an AutoML-driven world include:

• Data engineers focused on maintaining data/feature pipelines and lineage
• ML engineers specializing in model deployment, serving, and monitoring infrastructure
• "Analytics Translators" who can frame business requirements and oversee end-to-end AutoML projects
• Governance and audit experts to define responsible AI principles and practices

While AutoML is lowering the barriers to entry for building models, there is still a strong need for human-in-the-loop oversight to ensure that ML projects are technically sound, aligned with business objectives, and ethical.

Conclusion

Google Cloud AutoML is a powerful tool for accelerating the development of custom ML models without the need for deep data science expertise. By leveraging transfer learning and Neural Architecture Search under the hood, AutoML can often match or exceed the accuracy of hand-tuned models at a fraction of the engineering cost.

As we‘ve seen in examples across e-commerce, agriculture, life sciences and more, AutoML is unlocking the potential of ML for organizations of all sizes. It‘s enabling lean data teams to punch above their weight and rapidly prototype high-value use cases.

But AutoML is just one piece of the larger puzzle of making ML a mature, industrialized practice. MLOps platforms and practices are also needed to automate the end-to-end lifecycle of data prep, model serving, monitoring, and governance.

The combination of AutoML and MLOps is a major step towards making the transformative potential of AI accessible to every organization. As the famous computer scientist Andrew Ng has said, "I hope for a future where hundreds of thousands of companies are using not just predictive models, but optimization models and AI to help them figure out how to better serve customers, to better control our power plants, and do many other useful things."

With tools like Google Cloud AutoML, that future is fast becoming a reality. For companies looking to get started with ML without investing in an army of data scientists, AutoML offers a compelling jumpstart. It may not solve every ML challenge out of the box, but it‘s an invaluable complement to in-house ML expertise – and most importantly, a way to quickly prove the value of ML for your most pressing business objectives.

References

1. Notes from the AI frontier: Modeling the impact of AI on the world economy
2. 2020 State of Enterprise Machine Learning
3. Cloud AutoML: Making AI accessible to every business
4. AutoML-Zero: Evolving Code that Learns
5. Recognizing Bowl of Noodles using AutoML Vision
6. A Practical Guide to Image Classification with AutoML on Google Cloud
7. How we‘re using AutoML to detect Cassava diseases
8. The Cancer Imaging Archive (TCIA)
9. Using AutoML to Analyze The Cancer Genome Atlas Dataset in the Cloud
10. MLOps: Continuous delivery and automation pipelines in machine learning