MIT Gives Computers the Capability to Predict the Future with Deep Learning

February 06, 2017 by Robin Mitchell

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory have created an algorithm which significantly improves predictive ability.

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory have created an algorithm which significantly improves predictive ability.

An important trait that separates humans from other animals is our ability of prediction. Although some animals appear to have predictive abilities, such as hibernation, weather changes, and pack hunting, the human ability to predict is much more advanced. While the capabilities of animal and human prediction is far and varied one point is clear: the ability to predict is important! 

The Power of Prediction

Prediction helps us to anticipate dangerous hurricane weather patterns, determine if a farm crop will be plentiful, know whether a building on fire is about to collapse, and more. Billions are spent around the world on supercomputers which help to crunch numbers to make predictions as accurate as possible.

For example, Numerical Weather Prediction involves taking data such as barometric pressure, temperature, wind speed, and any other reading you can think of from the environment to produce a list of expected weather pattern results. The final weather prediction can then be made by meteorologists from the processed data helping to warm people of potential droughts, snow, and storms. Prediction of weather is helped drastically by computers performing the complex processing of gathered results. So what about a scenario that involves more intuition than math?


Computer models help to predict weather patterns - Image Courtesy Wikipedia [public domain]


One example of intuitive predictive ability would be anticipating other drivers on the road. Most drivers can report a time when they “felt” odd about another car and ended up either giving the car space or constantly keeping an eye on that driver. This type of prediction is something that computers seriously lack for several reasons. Firstly, computers are not adaptive learning machines and can only be told what to do via programming.

Currently, learning capabilities are in the form of neural networks which still limit a computer’s ability to behave like humans (mainly due to the sheer number of neurons, connections, and programmability of the brain). However, a team of researchers from MIT may have just changed the field of predictive ability.

Read More

More Complex Predictions

Imagine a computer that could predict the next action a human being will take. This is exactly what a research team from MIT have accomplished where a computer is shown millions of situations and uses these “experiences” to determine what will happen next. One example predicts whether two people will shake hands, hug, or kiss. The algorithm analyzes the meeting of two people and feeds the information into a neural network.

When the algorithm is exposed to a new situation the neural network can accurately predict what will happen next (however, it does make mistakes from time to time). Not only can the neural network predict the action, but the software also creates a 1.5-second video on what it thinks will happen.


Video Courtesy of MIT CSAIL

The still image which the algorithm receives is firstly split into two main parts, the background and then foreground. Then it further splits the foreground into the foreground and a mask. From there, it will attempt to make determinations on what the image will do and thus make the prediction. The system viewed 2 million online videos to observe how situations unfolded (for example, how a train moves past a station).


The predictive video generation system - Image Courtesy of Carl Vondrick/MIT CSAIL


Read More

Potential Applications

A computer that can predict 1.5 seconds into the future may not sound impressive but considering how quickly hazards can turn into dangers, 1.5 seconds may be all that is needed. Imagine a self-driving car that could constantly be watching out for danger, not by measuring speeds and detecting objects, but instead by looking at the road and trying to predict what other cars will do and what the environment is up to.

Such neural networks could also be implemented in the police force and security systems in general where the software could predict the move of a potential criminal. Imagine a security camera that could predict if an individual has an intention to commit robbery. The automated system could alert the police before the crime even took place. If the predictive capability could look far enough into the future, the police may only need to be sent as a deterrent instead of making any arrests (assuming the criminal offense had not taken place just yet).

But what about more controversial applications? Would such technology pose a threat to security? Could it violate people’s rights? Will Minority Report become a reality where people are arrested before they commit any crime?

Just like most technologies, there will always be a negative impact because some individual or organization has malicious intentions. But considering the fact that the prediction is 1.5 seconds, such prediction algorithms are more likely to save lives being utilized in accident avoidance and danger prevention.