Researchers at the University of Bern in Germany have designed a software that can watch videos and predict events that will occur in the next 5 minutes with an accuracy rate between 15% and 40%. The researchers say that for long videos with a large number of different actions, their method can also accurately predict the future, and even handle noisy or erroneous input information.
From the beginning of millions of years ago, the first hominids looked up at the stars and became curious about the universe. Our human understanding and control of the world around us has reached a fairly skilled level.
We can already fly at supersonic speeds, can transgene, and treat cancer. However, there is one thing we have not done very well, that is, the perception of "time"-how to infer the future from the present, how to make the most of this speculation?
Looking now, the machine seems to help us do this.
Computer scientists at the University of Bonn in Germany have designed a software that can predict sequence events in the next 5 minutes with an accuracy rate between 15% and 40%.
Although numerically, this accuracy is not high, but researcher Juergen Gall said that this represents machine learning beyond single-step prediction (Single-step prediction), a step forward to a new field. Related research papers have been accepted by CVPR 2018.
Juergen Gall team presents their system for predicting the future
Beyond single step prediction
The research goal of Gall and his colleagues—teaching computers to predict the future—is not just now. In fact, this is a major topic in the field of machine learning and computer vision. A large number of researchers are engaged in related work.
However, what makes this work special is its method: so far, research in these areas has focused on the interpretation of current actions, or the prediction of expected next actions, which is the aforementioned "single-step prediction" .
Single-step prediction is to predict the results of the previous step in the future. This is the basis for predicting the future model, which comes down to a regression problem, input variables, and prediction results.
In making such predictions, the current research has achieved relatively good results. One example is that the Stanford University Wu Enda team designed a deep learning algorithm that achieved up to 90% accuracy in the prediction of hospice.
In their experiments, Stanford researchers used 2 million medical records to train the neural network they designed. Through these data, the network can discover patterns and laws that doctors cannot find, and use this as a basis for new patients. The situation (the mortality rate of patients in the next 3 to 12 months) is predicted.
The variables faced by this research are very complex and have achieved very high accuracy. The paper also won the Best Student Paper Award of the IEEE Bioinformatics and Biomedical Group. However, such an algorithm is based on the retrospective and only performs single-step prediction.
Multi-step prediction: using CNN and RNN to predict two future structures
In the latest research by Gall and his team, they proposed two methods to predict a large number of actions that will occur in the future for a long time. They trained a CNN and an RNN, and learned the “tags†of future videos based on the video content they had seen before.
"We show that even for long videos with a large number of different actions, our method can accurately predict the future and can even handle noisy or erroneous input information." The author wrote in the paper.
The legend below shows two methods designed by the researchers. First look at the design of RNN.
In the structure of the RNN system, the input is a sequence, and the network predicts the remaining length of the most recently observed action, and the label and length of the next action. By appending the prediction result to the original input, the next action segment can be predicted.
In the structure of CNN system, the input sequence and output sequence are converted into the form of matrix. Where C represents the number of classes and S corresponds to the number of video clips of a specific length. The binary value of the matrix represents the label of each video clip.
Let the machine predict the long-term future and hope to achieve real human-computer collaboration
In their experiments, Gall and his team used hours of video data to demonstrate different cooking actions (such as frying eggs, mixing salads, etc.), and showed only some of the actions to the software. The software predicts the next action step based on what has been learned. Through this method, Gall hopes that this field can take a step towards true human-computer symbiosis.
"[Industry] people talk about human and robot collaboration, but in the end there is still a separation between humans and machines; they don't really cooperate closely," Gall said.
Gall suggested that by using appropriate hardware, this software can help humans accomplish tasks by intuitively understanding tasks, and then help humans work in an industrial environment.
Gall said: "There are more and more seniors, and it is worth trying to promote this robot in the family to take care of the elderly," Gall said: "I believe that after 10 years, the service robot will be able to take care of the elderly at home."
According to a report from the Census Bureau, the number of Americans over 65 today is about 46 million, and it is expected to double by 2060. According to a 2014 report from the US Centers for Disease Control and Prevention (CDC), about 1.4 million of these elderly people will live in nursing homes. Japan has already explored the impact of using such software. For example, the seal-type machine used for healing, PARO, Softbank's companion Pepper, etc. In Japan, a quarter of the population is elderly.
With the advancement of this kind of technology, it may cause further differentiation between human generations-outsourcing love and care to a machine. For an immature industry, it is difficult to say where this path will lead, but the final decision is in the hands of developers, not in the hands of the software or robots they develop.
Vacuum Cleaner Bldc Dry Motor,Vacuum Cleaner Bldc Motor,Dry Bldc Motor For Vacuum Cleaner,Bldc Brushless Motor Vacuum Cleaner
Zhoushan Chenguang Electric Appliance Co., Ltd. , https://www.vacuum-cleaner-motors.com