Overview of ML with GCP

Thousands of Google Engineers and Google customers have completed training similar to the content presented in this specialization. The goal of this specialization is to empower the people who complete it to understand how to train, deploy, and serve Machine Learning models on Google Cloud. It is intended to provide a practical, real-world intro to machine learning. It will provide the information necessary to help people with a variety of backgrounds, including Python programmers, data engineers, and data scientists, to do machine learning and build production machine learning models.

Agenda

Why Machine Learning?

1. How Google does Machine Learning describes what Google means when it says they are "AI first," and provides a high-level overview of Machine Learning strategy. It ends with a discussion of how ML at scale using server-less data processing.

Machine Learning with TensorFlow

2. Launching into Machine Learning elaborates the first aspect of machine learning: building a good dataset. Models that work well in experiments but then fail in production often fail due to a poor dataset.
3. Introduction to TensorFlow elaborates the second: building a good model using Google's open-source framework.

Note: Some of this content in this section was previously taught in the Coursera course “Serverless Machine Learning with TensorFlow”

Improving Machine Learning Accuracy

Both of these courses “fill up your ML toolchest.”

4. Feature Engineering describes the various ways machine learning models can be improved.
5. Art and Science of Machine Learning provides more theoretical background.

Note: Some of the content in this section was previously taught as part of the “Machine Learning Crash Course” that had been taught at some universities and is now hosted online.

Five other courses are also described, but these have been split off into their own specialization called “Advanced Machine Learning on Google Cloud Platform” that will be launching later this year.

ML at Scale - The first two of those courses, “Production ML Systems” and “End-to-End Lab on Structured ML Data” will discuss scaling machine learning systems.

Specialized ML Models - The final three courses delve back into machine learning theory, including “Image Classification Models,” “Sequence models,” and “Recommendation Systems.”

Why Learn ML from Google?

The following Google products use Machine Learning:

• Classify pictures in Google Photos
• Targeted ads in Adwords
• Smart reply in Inbox
• Recommendations for next video in YouTube
• Pedestrian detection for self-driving cars
• Span detection in Gmail

Google has been completely transformed by machine learning. In Q1 2012, Google had close to zero TensorFlow models in production. By Q4 2016, Google had produced over 4000 TensorFlow machine learning models.

Why Use Google Cloud?

Key lesson: To be successful at machine learning, your organization needs to think about serving ML predictions in addition to creating models. Operationalizing ML models is hard and must be a primary consideration. After all, the key reason your company invests in ML is to deliver recommendations and insights. So, to be good at ML, you need to be good at data engineering. “Lak” suggests data scientists take the data engineering Coursera specialization that Google offers.

“Operationalizing an ML model” is the process of taking a trained model and getting it to the point that it can serve out these predictions.

A lesson Google learned early on is that it needed to be able to process “Stream” and “Batch” data the same way. Google’s “Cloud Dataflow” (Apache Beam) enables this. Other products that form part of the Google Cloud Platform include the following:

• Cloud Pub/Sub, Cloud Storage, Cloud Dataflow, BigQuery, Bigtable, Cloud ML Engine, Cloud Datalab, Data Studio Dashboards/BI

The key services on GCP are serverless and they’re all managed infrastructure. By building data pipelines on Google Cloud, your organization can take advantage of the scalability and reliability of Google’s systems.