RE: Times today12 Aug 2020 06:19
“This is about the soul of data science,” Professor Pearl, who was not involved in the research, said. “Today’s machine learning is buried in this mentality that I call data centrality, as opposed to science centrality.” AI has found the “low-hanging fruit”, but expecting true intelligence to emerge is, he says, a little like “simulating evolution and expecting to get Einstein from an amoeba. It takes too long”.
To see why ignoring causality might be a problem, consider a computer program trained to look for the cause of flooding on roads. It might see that when streets are waterlogged lots of people also use umbrellas and conclude that umbrellas cause floods. In fact both have a deeper cause: rain.
Similarly, imagine an elderly smoker with chest pain, nausea and fatigue. Many people with those symptoms have emphysema, and a computer might conclude that this was the cause. A GP would know it is angina. The reason that lots of people have emphysema and those symptoms is not because one causes the other but because both have a deeper cause: smoking.
To train a computer not to fall into this trap, to help it spot causation, involves teaching it to consider “counterfactuals”. Would there still be floods without umbrellas? What if the patient did not have emphysema? Would the symptoms go away?
Using medical modelling, a little like the physics modelling used to create effects in video games, the computer is able to “imagine” what would happen if a disease — such as emphysema — was magically cured. Did the symptoms go too?
“If the symptoms did go away then we’d know that the causal path, the thing that generated the symptoms, was the disease,” Jonathan Richens, the lead author on the research, said. “In the case of emphysema, we know the symptoms won’t go away if it goes away because it doesn’t make you have chest pain or dizziness.”
Although it sounds obvious, this is not how computers generally think. But the new program does. “By imagining an alternate reality where the patient does not have the disease it can immediately get the problematic diagnosis.”
By testing it on 1,671 cases created by real GPs, Dr Richens, working with colleagues at UCL, found it significantly outperformed not only humans but also correlation-based AIs.
Professor Pearl said it was early days but he was “very hopeful they have broken a barrier here. With just a little bit of counterfactual you can get results which are unachievable with straight machine learning. I’m going to use it as a warning to machine-learning enthusiasts in the US — you're going to be made obsolete by companies in the UK.”
Sign In
Follow the stocks
