A guide to machine learning for the chronically curious: ML Explorer

I recently joined Google to edit this blog, and to explore the value of machine learning and big data in an intuitive and hands-on manner.

Over the past couple years, I've been fortunate to work with engineers who design and tune ML algorithms, and I’ve even trained my own models on a couple occasions. But since joining Google, I've been truly humbled by the techniques, code, and expertise of the software engineers, product managers, customer engineers, solutions architects, and developer advocates within Google Cloud. Not to mention the venerable researchers who sit on DeepMind and all Google AI teams. Some of the most capable minds in the world dedicate every working moment to machine learning: the art and science enabling computers to make increasingly sophisticated analyses.

Advanced machine learning techniques — deep neural networks to start, but also adversarial networks, long-short-term memory networks (LSTMs), and other architectures — help us craft automatic replies, recognize images, recognize speech, and a whole host of other applications. We’d like to share some of those techniques with you, because we think they’re amazing. And they might revolutionize your business, or perhaps just help your company compete better. A number of our customers have built in-house recommendation engines, quality control systems, and sales prediction tools with our products.

If you have some time to set aside for learning, I’d like to serve as a tour guide. Whether you’re a developer, IT strategist, or even a curious hobbyist, there’s plenty to discover by applying ML to your pre-existing or real-time datasets. But more specifically, I’m writing for data scientists, or for application developers who aspire to become data scientists.

Don’t worry if you don’t exactly fit these categories, though: ML happens to be fun, mostly because you can ask Google Cloud to execute ML-driven tasks that computers were previously and traditionally incapable of doing. We encourage you to try out any and all ML workloads on our cloud, and see what works for you or your business.

First, let’s dig in through an example you may be familiar with from the world of consumer electronics, Google Assistant. Although image recognition tasks such as ImageNet largely predated applications of deep neural nets on speech recognition, many consumers (and perhaps even more so after this past holiday season) are becoming familiar with AI in an increasingly conversational fashion.

Let’s set up an AIY device (RaspberryPi-based) to talk to a new Project on Google Cloud Platform. Using two additional components called Chatbase and Dialogflow Enterprise Edition, you’ll get to build your own chatbot to interact with you through the AIY device. You’ll prototype and deploy on GCP. Are you ready?

If you’re wondering what content will follow, here's a look ahead:

• Start with Assistant API via AIY Voice or the Actions Simulator site, then switch to Speech API (lower level, more customizable).
• Add on Chatbase and Dialogflow Enterprise Edition for customizable interactivity, and see the results and queries in Console.
• We’ll show how to train a generic Estimator model on a public dataset.
• We’ll then head back to vision applications by showing you how to train and deploy inception-v3 in TensorFlow on Cloud ML Engine.
• We’ll look at how a from-scratch trained solution compares to the managed Cloud Vision API.
• For a bonus, we’ll see how Google Brain trains robots in a simulation using PyBullet.