There’s now a computer that can predict the future, but only if it’s about making salad
Imagine what we could do with the power to look into the future? Would you look ahead to the winner of the World Cup and put down a big fat bet to reap the rewards? Perhaps you’d the power for good, stopping crimes Minority Report precog style. Whatever you’d decide to do, that possibility may not be all that far away anymore as a set of developers at the University of Bonn have given a computer program the power of foresight.
Well, actually, before we get too carried away here, the team hasn’t actually created a time-travelling computer. Instead they’ve developed software that’s capable of predicting, with great accuracy, how events will unfold before they actually happen. Well, salad-related events anyway.
So the software isn’t tearing a hole through space and time to peer into a single moment in time, then. It’s really just a highly-accurate prediction modeller based on the analysis of past information. That doesn’t make it any less impressive though. As Professor Dr Jürgen Gall explains, such a technique allows a computer “to predict the timing and duration of activities – minutes or even hours before they happen.”
One use case Gall’s team sees for the technology is in a robotic assistant. Humans can understand and anticipate the actions of others, providing support when it’s needed. Robots, however, struggle to do this. Now though, with the breakthrough at the Institute of Computer Science at the University of Bonn, computers can begin to estimate the timing and duration of future activities for periods of several minutes, understanding when they’ll next need to get involved to provide assistance ahead of a task being completed.
The software was trained on over four hours of salad videos, where different performers prepared different salads. Each recording lasted around six minutes, containing around 20 different actions. The computer was also shown, during the video, a time frame around what the action was, when it started and how long it took.
By watching these videos, the computer was able to learn which actions typically follow another and for how long they generally last. There needs to be room for flexibility though, different chefs do tasks in different orders or at different speeds and the computer needs to be able to understand their demands.
READ NEXT: This AI can see through walls
To test the software, the team threw a set of videos at it that it had never seen before. They were also about preparing salads, but this time around the computer only knew what was going to happen in the first 20% to 30% of the video – it then had to predict what was going to occur and when afterwards.
“Accuracy was over 40% for short forecast periods,” Gall explains, “but then dropped the more the algorithm had to look into the future.”
Activities that stretch further than three minutes into the future really tested the system, but it still gave back the correct prediction in 15% of the cases. Interestingly, being “correct” was only the case if the computer managed to predict both the activity and its timing, rather than just one or the other.
The team also realised that the computer wasn’t very good at working out what was happening in a video with no contextual information about how it started. However, that’s a problem that can be resolved by scrubbing the data and helping provide some learning context to the program itself. According to Gall, the data is too “noisy” for it to work well. “Our process does work with it, but unfortunately nowhere near as well,” he explains.
So, if you can’t wait for a computer to look into the future and help tell you what to do next with a salad, you’re in luck. If you’d like it to give you the winning lottery numbers, sorry pal, you’re going to have to wait a little bit longer for that to happen.