AI has advanced to another hallmark approaching human likeness: predict the movement of objects. This is thanks to a new prediction algorithm described in a published paper entitled "[Anticipating Visual Representations from Unlabeled Video](http://web.mit.edu/vondrick/prediction.pdf)".

<center><img src="http://www.steemimg.com/images/2016/12/05/prediction2fbede.jpg" alt="prediction2fbede.jpg" border="0"></center>
<center><em>[source](https://www.re-work.co/blog/machine-intelligence-carl-vondrick-mit-computer-vision-predictive-intelligence)</em></center>

The developers, Carl Vondrick, Hamed Persiavash and Antonio Torralba, developed a way to train the deep networks to predict visual representation of images in the future. They tested their prediction algorithm to [anticipate objects both one second, and five seconds](http://web.mit.edu/vondrick/prediction.pdf), into the future.

<center><img src="http://www.steemimg.com/images/2016/12/05/prediction31d7bc.jpg" alt="prediction31d7bc.jpg" border="0"></center>
<center><em>[source](http://www.dailymail.co.uk/sciencetech/article-3652736/The-AI-predict-future-relationships-Software-binge-watched-Office-Scrubs-learn-humans-interact.html)</em></center>

The AI was shown unlabeled images to generate a future prediction. Unlabeled means no tags or information associated. The computer then tries to predict what will happen next based on the objects it can detect in the image and what it perceives to be happening.

<center><img src="http://www.steemimg.com/images/2016/12/05/prediction104590.jpg" alt="prediction104590.jpg" border="0"></center>
<center><em>[source](http://web.mit.edu/vondrick/prediction.pdf)</em></center>

The still image is presented and the computer tries to generate a mini movie based on that still image. The program has a before action image, and then generates a forecasted 1 second and 5 second ahead images.

When it is uncertain, the program comes up with multiple forecasting predictions:

> "Multiple Predictions: Given an input frame (left), our model predicts multiple representations in the future that can each be classified into actions (middle). When the future is uncertain, each network can predict a different representation, allowing for multiple action forecasts. To obtain the most likely future action, we can marginalize the distributions from each network (right)."

<center><img src="http://www.steemimg.com/images/2016/12/05/prediction47465d.jpg" alt="prediction47465d.jpg" border="0"></center>
<center><em>[source](http://web.mit.edu/vondrick/prediction.pdf)</em></center>

To get better videos, [adversarial networks](https://en.wikipedia.org/wiki/Generative_adversarial_networks) were used. One neural network generates video while another looks at it to judge if it's fake or real. One network tries to become better at making fake video, and the other tries to become better at judging. 

The technology can only last for just over a second, with a low-resolution 32 frame total output being generated. Sometimes it predicts accurately, others not. The further ahead it tried to predict, the more blurry the images generated.

A frame of a train, can indicate which direction a train is heading, but it still doesn't have common sense without specific rulesets configured. The program doesn't understand that eventually the moving train will move out of the still frame. The program doesn't understand how the world works in general, no common sense. 2 million videos are being used to teach the program to predict.

Here is a short video explanation as well:

### Action-Prediction Algorithms
https://www.youtube.com/watch?v=AR3hY9iB5-I

<br/>

## Why is understanding how to predict the future important? 

Because when we can anticipate what will happen, where something will be, what someone will do, then we ourselves can plan ahead based on the input from our environment that changes. Rather than rely on real-time input before making a move now, the AI will be able to act more like humans, and plan ahead based on where something will likely.

This can help in developing robotic assistant technologies, such as health care for someone who is about to fall, catching them as a human would, or a self-driving car that predicts an oncoming care accident and avoids it. 

A robot that operates in real-time alone, is not able to function alone and independently in a human world. Being able to anticipate problems before they happen is a real-world requirement. Failure to operate in the future dimension makes someone a sitting duck waiting for things to affect them. This is not an optimal survival feature, and certainly being limited to this capacity also limits the extent of interoperability the technology can have with our human world.

Computer vision and machine learning are being combined more and more to advance AI development. I have covered aspects of this in the past with respect to learning facial recognition and learning sounds that objects make, all using video or images to learn and develop "thinking" algorithms. Despite no one being able to predict what will happen in the future, not even an AI, the AI may be able to help show us alternative futures we may not have thought of.

----
#### Thank you for your time and attention! I appreciate the knowledge reaching more people. Take care. Peace.

---
References:
[Anticipating Visual Representations from Unlabeled Video](http://web.mit.edu/vondrick/prediction.pdf)
[Carl Vondrick](http://web.mit.edu/vondrick/)
[Artificial Intelligence Algorithm Predicts the Future](http://www.newsweek.com/artificial-intelligence-algorithm-predicts-future-473118)
[AI learns to predict the future by watching 2 million videos](https://www.newscientist.com/article/mg23231020-100-ai-learns-to-predict-the-future-by-watching-2-million-videos/)
[New AI can predict when two people will kiss](http://money.cnn.com/2016/06/21/technology/ai-kissing/)
[MIT Creates AI Able to See Two Seconds Into the Future](http://www.dailygalaxy.com/my_weblog/2016/11/mit-creates-ai-that-is-able-to-see-two-seconds-into-the-future-on-monday-the-massachusetts-institute-of-technology-announce.html)
[Predicting Future Human Behavior With Deep Learning](https://www.re-work.co/blog/machine-intelligence-carl-vondrick-mit-computer-vision-predictive-intelligence)
[Teaching machines to predict the future](http://news.mit.edu/2016/teaching-machines-to-predict-the-future-0621)

---
If you appreciate and value the content, please consider:
Upvoting <img src="https://www.steemimg.com/images/2016/08/30/upvote91a69.png" alt="upvote91a69.png" /> ,&nbsp;&nbsp;&nbsp;&nbsp;Sharing <img src="https://www.steemimg.com/images/2016/08/30/share2195b.png" alt="share2195b.png" /> **and** &nbsp;&nbsp;Reblogging <img src="https://www.steemimg.com/images/2016/09/19/reblog33b5f.png" alt="reblog33b5f.png" /> **below**.

[![Follow](https://www.steemimg.com/images/2016/08/30/follow2be5e.png)](https://steemit.com/@krnel) me for more content to come!

---
@krnel
2016-12-05, 5pm