Despite becoming a something of buzzword in recent years, the fact remains that any company serious about its future is looking into or trying to integrate machine learning (ML) into its operations. In a presentation by Google solutions architect Robert Saxby at the AI & Big Data Innovation Summit in London, Saxby delved into how Google started to create ML models to improve its energy efficiency and why it began to focus on ML automation.
"We are in the business of data centers," Saxby explained. "YouTube alone streams one billion hours of video every day." With these amounts of data being stored, shared and accessed regularly by Google and its subsidiaries such as YouTube, Google has become one of the planet's top five largest data center companies. However, with all this data comes the need for larger data centers, and with larger data centers come significantly higher power demands.
Therefore, in 2016, Google decided to create an AI-powered recommendation model which could suggest adjustments to data center cooling systems in an effort to save energy. Google harnessed the ample amount of data it had already collected from its data centers and used it to train its new recommendation model.
And it worked! The model was able to reduce its data centers' energy consumption by 40% almost immediately. However, it remained a manual recommendation engine, which meant it still needed a person behind the wheel making the final decision based on the model's suggestion. And while the model performed great when there was someone there to make that choice, if someone was not, then the model was essentially useless.
So, in 2018, Google decided to automate the process completely, moving away from a human-implemented recommendation and placing the AI system in direct control of the cooling system. Once it had carried out these changes, the company witnessed a further 30% improvement in data center energy consumption. Google already had a pipeline feeding the model with raw data, with the AI running the system and the new data constantly retraining the model. The model continued to improve itself, hence, the data centers' energy consumption has continued to fall as a result.
In considering the significance these types of models could have on other industries, Saxby said Google began asking itself "let's see how we can apply this to other systems".
Saxby highlighted a number of the ways and benefits of ad-hoc reporting and analysis, but illuminated the biggest problem with smaller companies trying to create their own models from scratch – namely, its immense time and labor intensiveness. This is where systems such as Google's AutoML comes into play.
While every company can and should be collecting as much raw data as possible, the operationalization of this data can be difficult as not every enterprise has the resources or expertise Google has to train models from the ground up. However, Saxby explained that due to Google's range, its models are quite generic, which allows for them to be built upon. So, by selecting one of Google's generic ML models, and inputting your own data and parameters, you can retrain AutoML into a custom model useful to you alone.
Saxby uses the example of sentiment analysis to explore this point. If a retail company wanted to be able to respond to negative feedback posted online quickly, it would need a model able to analyze text and pick up on the negative comments. For a retail company or other similar businesses, a generic sentiment analysis model would suffice. However, if a firm belonged to a niche field or industry where the terms used may not be the ones a generic sentiment analysis model may have been trained on, one would require a much more specific model.
With a platform like AutoML, you can combine your data with Google's API and create your very own model. The leveraging of these kinds of models for custom use is what Saxby referred to as the "democratization of ML for all audiences". It means that not only can big companies with the money and resources create their own models, but so to can the little guy, bringing in all the saving and insights that come with it.
"Think about how you collect data, how you process data, what you can do with it and the pipeline end-to-end," concludes Saxby. "You don’t have to do it all at once but there are ways to get there (to your own models) quicker by using well-founded ML models and incorporating your data into it."
Find out more HERE