Machine that anticipates your movements several minutes in advance underway

An intelligent machine capable of anticipating your next move minutes in advance sounds like the stuff of nightmares – but is now a reality. Researchers have taught an AI to recognise patterns in people’s actions, allowing it to accurately predict the next move in a sequence minutes in advance. The software, which was built by […]

An intelligent machine capable of anticipating your next move minutes in advance sounds like the stuff of nightmares – but is now a reality.

Researchers have taught an AI to recognise patterns in people’s actions, allowing it to accurately predict the next move in a sequence minutes in advance. The software, which was built by a team at the University of Bonn in Germany, was taught to anticipate actions by watching hours of cooking videos.

Dr Jürgen Gall believes the intelligent software will eventually be able to prophesize your actions ‘hours before they happen’.

If the team manages to fine-tune the algorithm to be able to anticipate actions that far in advance, it's possible to imagine a slew of real-world application, from home automation gadgets, to Big Brother-esque surveillance.

Researchers told the AI what was happening in the video for the first 20 per cent of the clip –and then asked the algorithm to predict the next action before it took place To teach the AI to accurately predict actions before they take place, Dr Jürgen Gall and his team focused on cooking videos.

Using pre-recorded videos of people preparing a meal, the researchers were able to teach the machine to recognise each action being performed on-screen, including cutting tomatoes, adding salt and flipping a pancake. In total, some 40 videos were used to teach the AI. Each of these recordings was around six minutes long and contained some 20 different actions.

After four hours, the algorithm was able to recognise the sequence of events needed to prepare a dish, which is far from trivial given the variety in approaches and recipes in the pre- recorded clips.

‘Then we tested how successful the learning process was,’ explained Dr. Jürgen Gall. ‘For this we confronted the software with videos that it had not seen before.’ Like before, the machine was told what was happening in the video for the first 20 or 30 per cent of the clip. The algorithm was then asked to predict the next action before it took place on-screen. The machine flagged-up its 'observation' before the action, drawing on its knowledge of the recipe and its understanding of how similar sequences have played out before.

The AI was able to correctly anticipate actions in the near future with surprising accuracy. Dr Gall said: ‘Accuracy was over 40 percent for short forecast periods, but then dropped the more the algorithm had to look into the future.’

For activities which were more than three minutes in the future, the algorithm was only able to accurately predict the outcome in 15 per cent of cases. Researchers only considered the prediction correct if both the activity and its timing were correct.

Gall and his colleagues want the study, which will be presented at the Conference on Computer Vision and Pattern Recognition in Salt Lake City on June 19, to be understood as a first step into the new field of activity prediction.

According the researchers, the algorithm performed noticeably worse when it was forced to recognise the actions in the first part of the video, instead of being told.

The project was spearheaded by Dr. Jürgen Gall who worked alongside Yazan Abu Farha to teach the AI to anticipate actions ahead of time. Both Dr Gall and Abu Farha created the machine at the Institute of Computer Science at the University of Bonn, Germany