Training machine learning models to deliver business value

by

24 July 2018
Google - Valliappa Lakshmanan

Businesses looking to derive value from artificial intelligence (AI) should focus on applying machine learning to growing volumes of structured data in scenarios where repeated decisions have to be made, said Valliappa Lakshmanan, Technical Lead for Machine Learning and Big Data at Google Cloud.

Speaking at the Google Next conference on 24 July, Lakshmanan described machine learning as “a way to use standard algorithms to derive predictive insights from data and make repeated decisions”.

”Don’t apply machine learning if you need to make decision once a year or once a month. That is not a machine learning use case,” he said. “A machine learning use case is a decision that you need to make over and over again.”

An example he gave was the decisions that have to be made to persuade online shoppers not to abandon their shopping carts, which would apply to every visitor to the site.

In his presentation on “Leveraging AI on the cloud to transform your business”, another insight that Lakshamanan shared was that “it’s not who has the best algorithm who wins; it’s who has the most data”.

“When the machine learning model is trained on more data and better data, the data will control the quality of the result,” he said.

Conventional methods of managing and analysing data tend to involve filtering down data and aggregating it. His advice: “Go back and train your model not on filtered or aggregated data but on raw data. Machine learning is about doing things on as much data as you can. Don’t aggregate or filter things too early.”

“Deep learning only works because data sets are large. Each time a data set doubles, the error rate drops linearly,” he said. “We are talking about thousand times, million times more data.”

The other "scary" part of curve is that the compute capability needed for state-of-the-art machine learning models – petaflops per day of training – is also growing at an exponential rate. From Alexnet in 2012 (a neural network for image classification) to AlphaGo Zero in 2017 (a version of AlphaGo which does not require inputs from human games), there was a 300,000x increase in compute power involved, he said.

So what happens when businesses collect petabytes and exabytes of data and need that magnitude of compute power?

Businesses will not be able to deliver value from data if they are focused on building and maintaining the infrastructure needed for typical big data processing, said Lakshamanan.

Google Cloud’s answer to this is a server-less approach to data analytics, data processing and machine learning - having these capabilities delivered as a fully managed service instead.

It is also lowering the compute barrier with its Tensor Processing Units - custom chips that dramatically accelerate machine learning tasks and are easily access through the cloud.

At the Next conference, Google also announced the extension of AutoML to cover natural language and translation capabilities. AutoML is a suite of machine learning products that enables machine learning models to be trained to address business needs, even by people with limited machine learning expertise.

Fei Fei Li, Chief Scientist for Google AI, said AutoML extends machine learning models to businesses without them having to write any code. She gave the example of the first release, AutoML Vision, which provides cloud vision APIs for businesses to train image recognition models. For example, by uploading as few as 10 images, and by clicking a “Train” button, a business can start training the model to recognise objects and, more importantly, to generate predictions on images that model had not seen before.

Google also announced updates to its core machine learning APIs.

  • The Cloud Vision API now recognises handwriting, additional file types (PDF and TIFF) and product search, and can identify where an object is located within an image.
  • Cloud Text-to-Speech provides multilingual access to voices generated by DeepMind WaveNet technology and the ability to optimise for the type of speaker, and
  • Speech-to-Text now has the ability to identify what language is spoken as well as different speakers in a conversation.