bluemoon1981 -

HPE calls for bottom-up approach to AI

A bottom-up approach towards data modelling is needed to address the shortcomings of physical and theoretical models in artificial intelligence

From financial modelling to using physical models to ascertain the performance of an aircraft engine in turbulent conditions, the use of economic theory and physical laws to predict outcomes has served humanity well.

However, such top-down approaches of developing artificial intelligence (AI) models are not always effective, especially if humans are included in the equation.

Citing the example of models that failed to predict the irrationality of people in the derivatives market, Goh Eng Lim, chief technology officer for high performance computing (HPC) and AI at Hewlett Packard Enterprise, said a bottom-up approach towards AI modelling was needed.

Speaking at Supercomputing Asia 2018 organised by Singapore’s National Supercomputing Centre, Goh said this approach would be entirely based on historical data to predict future outcomes without using existing models grounded in theory. “This has already started in the fintech [financial technology] world, and we’ll start to see this HPC,” he said.

Goh tested this approach when he got his researchers to scour for historical weather data on the internet, and run the data through a machine learning algorithm to predict if it would rain at the San Francisco airport. This was done without performing time-consuming calculations on supercomputers using meteorological models.

After repeatedly tweaking the weightage of different data points to predict the probability of rain, the algorithm managed to achieve a relatively good accuracy of 82%. Although the use of a meteorological model could increase accuracy to 95%, Goh’s team could obtain a prediction within seconds rather than hours.

Goh, who worked with the US space agency to send a supercomputer to space, said this bottom-up approach based on inductive reasoning is useful in space research where less powerful computers onboard a space facility can be used to crunch data for quick answers. “The more accurate answers are going to come from Earth later,” he said.

Read more about high performance computing and AI

Some researchers are even combining data-driven models with deductive top-down theoretical ones in a bid to further improve accuracy and derive a more precise answer, said Goh.

Companies such as GE are already working to combine physical and data-driven machine learning models in its industrial systems. Joshua Bloom, vice-president of data and analytics at GE Digital, said at the Strata Data Conference in Singapore in December 2017 that doing so will help to improve physical models.

Bloom added that the company was also ensuring that the information generated from its models was trusted and understood by users. “Even if the models are correct, if people don’t trust and accept them, they are not going to wind up being used,” he said.

“In the end, we need to create not just better algorithms, but also make machine learning suggestions understood to people with domain expertise. We also need to build systems that take in feedback, and are cognizant of the user and the effects of a good and bad answer.”

Read more on Artificial intelligence, automation and robotics