Beware the black box

WHILE artificial intelligence and machine learning are playing increasing roles in mining going forward, experts warn against black box thinking.
Beware the black box Beware the black box Beware the black box Beware the black box Beware the black box

Penny Stewart at Future of Mining Australia 2019.

Petra Data Science managing director Dr Penny Stewart, who is speaking at the Future of Mining Australia 2020 event in Sydney, said bringing transparency to the area was something her company was working to do.

"If people don't understand what's going on under the hood, they don't trust it or have confidence in it," she said.

"It's all well and good to have these technologies and they all work well.

"The early adopters grab it and do well with it."

However, the problem with this black box approach arises after those early adopters move on.

The people who follow them can be less inclined to take on faith what the black box is spitting out at them than their predecessors.

They end up falling back on their old approaches and the gains that were made with artificial intelligence and machine learning are lost.

"We've had this experience with a number of companies where it worked really well," Stewart said.

"In one case a geo had really good success with it. He published papers on it.

"Then he left and the people who took over were more conservative. They stopped using it even though it had been working for them."

Data 61 senior research engineer Dr Dave Cole told the recent CSIRO Resources Innovation Showcase that it was important miners be critical of machine learning and artificial intelligence results.

"It you give me some data I can give you a model," he said.

"But you need to have humans doing some geological work in the region to drive that criticism."

Stewart puts it another way: be aware of the inputs going into the box to better understand the outcomes.

"We have tools called SHAP plots," she said.

"They show you what goes into the model. It gives you the relevant importance of the inputs to the model.

"Some things people think are really important aren't necessarily that important.

"These show what is really driving the optimisation of what they are trying to optimise.

"Sometimes there is a real ‘wow that's interesting' when you show the plots.

"Then there are some plots that show the relationship between different parts of the geology."

Stewart said that was important because engineers and geologists often tended to look at a one-way process.

"If you are trying to increase crusher throughput you think if there are harder rocks going through it slows the crusher down," she said.

"But it can be the softer rocks that slow things down."

It is in the right mix of hard and soft rocks that optimisation can be found.

"That's an example of the complexity machine learning can bring you," Stewart said.

"People know this exists. They've seen it empirically. But how do you build a model to capture this?"

Stewart said as the acceptance of machine learning and artificial intelligence grew in the mining space, so too did the need to make things more transparent.

 

"Geologists want to understand what they were being given," she said.

"Now as we're getting more maturity in the market and scaling up, a lot of work with our customers has been on this glass box approach.

"It also allows people to interrogate the models."

Dingo is another company that deals with masses of data to help miners optimise outcomes - although its approach is much more focused on improving equipment uptime.

Its executive chairman Paul Higgins said the market had reached the point where a lot of the customers were over the hype but were wary of some of the big investments that had been in made in the past that had gone nowhere.

"Dingo's approach is teaming with customers that have a particular problem and it's important enough to them to give us some data," Higgins said.

"Our domain experts and their domain experts come up with something that delivers value.

"We do see that we need to bring the customer along on the journey.

"Otherwise their default position is ‘at the end of the day if this stuff doesn't work, I'll have to go back to what I know. If you are asking me to do something different to that I need to understand it because there's a lot of risk'."

Another part of Petra's work is on putting its machine learning and artificial tools into the platforms miners already use.

"If engineers can use these tools in their normal workflow it makes it easier for them to use them," Stewart said.

"Last year in particular we spent a lot of time building partnerships to make it look like business as usual for them."
One example of that is the partnership Petra struck with Maptek to have its tools co-located in the Vulcan environment.

"In the processing plants the outputs from [Petra's] Maxta goes into Pie, the screens the operators use to control the plant," Stewart said.

"We can get our outputs to their control screens.

"The feedback we get is that it's getting increasingly difficult to get senior engineers in the mine planning area.

"If they have a tool that helps them do their work more quickly they embrace it because they are understaffed."