Artificial Intelligence going Rogue

Author: Shazin Sadakath


Artificial Intelligence dates back to 1940's with the inception of research into Neural Networks but it didn't get much traction because of the lack of computing power. But AI got a renaissance in the last few years specially after Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton from University of Toronto (research paper) managed to submit a successful Deep Learning Convolutional Neural Network (CNN) implementation to classify ImageNet dataset back in year 2012. 

Nowadays so many research and development is taking place on the subject of Artificial Intelligence and specially with multi layers of Neural Networks stacked together called Deep Learning. There are established commercial players like Amazon, Google, Microsoft offering Artificial Intelligence as a Service. And also frameworks like Tensorflow, Deep Learning for Java (DL4J) which enables to develop, train and use AI models easily. 

But Artificial Intelligence is all about letting the model to decide how to solve a problem it is provided with such as classification, generation, etc. This means there can be times when the AI system might try to cut corners to achieve its ultimate goal. 

For instance an AI system developed by Randlov and Alstrom to enable Reinforcement Learning and Shaping to Drive a Bicycle. In order to speed up the process they positively rewarded the agent when its made progress towards its goal. The AI system cut corners and learned to ride in tiny circles near the start state because no penalty was incurred from riding away from the goal.

A similar problem occurred with a soccer-playing robot being trained by David Andre and Astro Teller. Because possession in soccer is important, they provided a reward for touching the ball. The agent learned a policy whereby it remained next to the ball and “vibrated,” touching the ball as frequently as possible. 

A user by the name Victoria as listed many of these so called "Specification Gaming" in one of her google document.

As with hollywood sinister AI systems such as the classic HAL-9000 from the movie 2001: A Space Odyssey to SkyNet from the Terminator series seems not that far away. Only time can tell. 



Tags: ArtificialIntelligence Sinister Rogue AI
Views: 189
Register for more exciting articles

Comments

Please login or register to post a comment.


There are currently no comments.