People in technology know there’s more and more data being created, but artificial intelligence startup Numenta is tackling a slightly different problem: the speed at which data is produced.
The Silicon Valley company, founded by mobile computing pioneer Jeff Hawkins, last week said energy-efficiency company EnerNoc is among a small group of businesses now testing its software.
Its product, called Grok, is designed to process data to yield some sort of prediction or insight. The software is modeled on how the human brain processes data and is particularly well suited for handling streams of information, such as data from a sensor on regular intervals, the company says.
EnerNoc is using Grok to earn more revenue from its demand response customers who agree to reduce electricity. Grid operators pay EnerNoc for aggregating many of these energy reductions at buildings to maintain a steady grid frequency. Grok studies data from utility meters to predict how reliably an EnerNoc customer can be counted on to lower electricity use in the next two hours.
Another unnamed customer is a wind farm operator in Europe, which needs to make sense of data produced by as many as 34 sensors on 800 turbines, according to Numenta. Grok has detected patterns in gear box temperatures to flag potential problems and send out a maintenance person, a costly task for offshore wind turbines.
Numenta is also working with a mobile ad networks to maximize revenue from different sources of its ad inventory and with hedge fund managers to predict stock market prices. Another application of predictive analytics software is to recognize a fraudulent transaction immediately.
But energy has emerged as one of the most compelling uses of its technology because buildings and energy equipment are increasingly being instrumented, such as a wind or gas turbine. “The challenge you find in energy is that there’s just an explosion of data sources,” says CEO Rami Branitzky.
In conventional analytics systems, data is collected and experts design a model for organizing it and studying it. Grok doesn’t move the data into a separate system for analyzing it later. Instead, the software can take a stream of information and create a model that can identify patterns and relevant properties automatically.
The cloud-based software could, for example, predict energy use in a building based on previous patterns or take diagnostic information from a car’s engine and recognize an anomaly. One of the main advantages of the software is that it relieves the bottleneck of finding data scientists who create the models that make data useful, says Branitzky. The company charges customers based on the number of models the software it creates, which can be made from multiple data streams.
The software was developed to mimic the workings of the neocortex, the outer layer of the brain responsible for planning, speech, and movement. (See, The Brainy Learning Algorithms of Numenta.) It’s an attempt to improve on machine learning, a field of artificial intelligence where computers can act without having explicit instructions written as programs. (For a deep dive on Grok, here’s a technical presentation from Hawkins.)
In the software world, we’ve heard about the great promise of artificial intelligence for years and Hawkins has been working on Numenta since at least 2005. So there’s reason to be skeptical of Numenta’s performance claims.
What has changed since Hawkins first started the company is the amount of data that can be analyzed, particularly devices other than traditional computers, such as sensors on a wind turbine. Numenta now needs to find the applications where its software outshines traditional predictive tools and find customers willing to pay for analytics on the fly.
When designing an embedded system choosing which tools to use often comes down to building a custom solution or buying off-the-shelf tools.