In news that in no way will lead to a real-life AI version of Minority Report, where robots can predict your future crimes, researchers have developed a machine-learning algorithm that can predict future behavior with astonishing accuracy.
This newly developed AI is less sinister than the Spielberg film, though we admit it's still bordering on the mildly creepy end of the future-tech spectrum. Created by researchers from the University of Bonn in Germany, the machine-learning algorithm is able to predict the duration of actions of human subjects, as well as what will happen next, though no crimes as yet.
In fact, the AI is able to predict with over 40 percent accuracy what will happen in the next few minutes, with accuracy dropping down to 15 percent when asked to forecast events more than 3 minutes into the future.
How did they achieve this feat? Well, first the researchers showed their algorithm videos of people performing a task with many different variables – in this case, preparing a salad.
The AI "watched" 40 videos of humans preparing a salad, with each 6-minute video containing around 20 different actions. There were variations between how each human made the salad ie the order of ingredients used, as well as the type and quantity, and the way they were prepared.
The poor AI watched 4 hours of this mundane footage, in order to learn which tasks typically followed what, and how long each individual task took.
“Then we tested how successful the learning process was”, Dr. Jürgen Gall explained in a statement. “For this, we confronted the software with videos that it had not seen before.”
The AI was shown yet more videos of humans prepping salads – but only the first 20 or 30 percent. The algorithm was then asked to predict what would happen in the rest of the video. A prediction was only considered correct if the activity and duration of the activity were correctly guessed in advance.
“Accuracy was over 40 percent for short forecast periods, but then dropped the more the algorithm had to look into the future,” Dr Gall said.
The performance dropped off when asked to predict more than 3 minutes into the future, though the team wants to work on that.
“We want to predict the timing and duration of activities – minutes or even hours before they happen."
The researchers, who will present their findings at the Computer and Pattern Recognition Conference in the next few days, think there are practical applications for their study that could be used in the near future. For instance, an oven could know when to preheat ready for its chef, or a robot vacuum cleaner could predict how long you'll be in another room and get the cleaning done while you're out of the way.
Fortunately, no word yet on when your toaster can predict your crime spree.
