19/06/2018
Teaching computers to plan for the future
As humans, we’ve gotten pretty good at shaping the world around us. We can choose the molecular design of our fruits and vegetables, travel faster and farther and stave off life-threatening diseases with personalized medical care. However, what continues to elude our molding grasp is the airy notion of “time” — how to see further than our present moment, and ultimately how to make the most of it. As it turns out, robots might be the ones that can answer this question.
Computer scientists from the University of Bonn in Germany wrote this week that they were able to design a software that could predict a sequence of events up to five minutes in the future with accuracy between 15 and 40 percent. These values might not seem like much on paper, but researcher Dr. Juergen Gall says it represents a step toward a new area of machine learning that goes beyond single-step prediction.
Although Gall’s goal of teaching a system how to understand a sequence of events is not new (after all, this is a primary focus of the fields of machine learning and computer vision), it is unique in its approach. Thus far, research in these fields has focused on the interpretation of a current action or the prediction of an anticipated next action. This was seen recently in the news when a paper from Stanford AI researchers reported designing an algorithm that could achieve up to 90 percent accuracy in its predictions regarding end-of-life care.
When researchers provided the algorithm with data from more than two million palliative-care patient records, it was able to analyze patterns in the data and predict when the patient would pass with high levels of accuracy. However, unlike Gall’s research, this algorithm focused on a retrospective, single prediction.
Accuracy itself is a contested question in the field of machine learning. While it appears impressive on paper to report accuracies ranging upwards of 90 percent, there is debate about the over-inflation of these values through cherry-picking “successful” data in a process called p-hacking.
In their experiment, Gall and his team used hours of video data demonstrating different cooking actions (e.g. frying an egg or tossing a salad) and presented the software with only portions of the action and tasked it with predicting the remaining sequence based on what it had “learned.” Through their approach, Gall hopes the field can take a step closer to true human-machine symbiosis.
“[In the industry] people talk about human robot collaboration but in the end there’s still a separation; they’re not really working close together,” says Gall.
Instead of only reacting or anticipating, Gall proposes that, with a proper hardware body, this software could help human workers in industrial settings by intuitively knowing the task and helping them complete it. Even more, Gall sees a purpose for this technology in a domestic setting, as well.
