How modeling user intent could revolutionize what we can do with AI
Technology users are bad at telling artificial intelligence what they really want, and this drastically limits the scope of its use. Through a unique approach, FCAI researchers hope to overcome this limitation – and fundamentally change how we develop AI.
Researchers from the probabilistic machine learning group at work. Image: Matti Ahlgren
Artificial intelligence has enormous promise for our future, but there is an essential problem that limits the extent of its usefulness: it cannot understand the intentions of us technology users.
The solution to this problem might sound like a no-brainer – just tell the AI what you’re aiming for so it can go about realizing it – but in practice, this is where much of AI use runs into a brick wall. More often than not, it turns out, we don’t really, in an exact enough way, know what we want.
“Though current AI learns from data and can even operate independently when given a precise goal, when users are not able to clearly formulate what they want, AI cannot help”, explains Samuel Kaski, the director of FCAI.
“In consequence, the amount of problems that artificial intelligence currently can solve is extremely limited.”
So there is tremendous potential yet unrealized in AI. How can we tap into it?
At FCAI, researchers are developing solutions that tackle precisely this issue. The ambition is to develop AI that helps uncertain users formulate goals. To do this, the AI would give users alternative choices and show them the outcomes of those choices, thereby helping the users select appropriate next steps as they go along.
For such a process to be possible, the AI needs to anticipate people’s actions and understand their intentions; in a sense, it has to have a model of the human mind.
To better understand how the human mind works, and to capture its essential aspects in a model, FCAI researchers combine AI, human-computer interaction, and cognitive science. (Read more about combining AI and cognitive science here.) In this work, FCAI is a pioneer. Typically, research on machine learning tends to leave the user out of the equation, whereas research on human-computer-interaction rarely involves AI.
“We consider this focus a major accomplishment of ours”, Kaski says, “and a central aspect that separates our research from that of others”.
How it got started, and where it’s going
The idea for this research focus emerged from a project on a new kind of search engine in the early 2010s. The goal was to develop an engine that successfully helps users who are not quite sure what they are looking for to better define their goals, and thereby find more relevant content. To do this, the search engine was constructed so that it interactively builds a model of user intent. This it uses to make keyword suggestions that augment the original search. (Read more and watch a video about newest results that have spurred from the project.)
“We realized that this same need for AI-assisted support on formulating user intent is relevant in problem-solving far beyond the original context”, Kaski says.
On the basis of this research project, the company Etsimo Healthcare – currently a partner of FCAI – was founded. The broader insights stemming from the project were formulated into the joint methodological goal that drives much of FCAI research today: The ambition to develop AI-assisted decision making, design, and modeling.
Currently, FCAI researchers approach this goal from a multitude of angles. The ambition for AI-assisted decision making can, for instance, be seen in the new, interactive technology developed by FCAI researchers that improves the accuracy of prognoses for medical treatments by setting up a dialogue between the prognostics application and medical doctors. This technology was first developed and tested in cancer research, but is applicable to other diseases as well. First results were published in 2018.
All in all, FCAI research hopes to initiate a paradigm shift which will massively expand the realm where AI can be of use – to fundamentally change the way AI is developed, and to broaden the scope of problems that it can help people solve.