Where Has Artificial Intelligence Gone So Far?

Computing

"I will destroy humans". This is how Sofia, a humanoid robot, answered open questions in an interview. This highlights the scary part of artificial intelligence, science fiction movies coming into reality and robots trying to take over the whole world.

Certainly, this is not the intention for artificial intelligence which is making major advances in the recent years. It seems that answers are coming to questions that we always asked about artificial intelligence. Will the machines be able to learn by themselves? Could they get rid of people and learn techniques and ways to solve complex games and life problems?

In the past years, several ways have been developed to collect data from human interactions with the surroundings. Then the machine processes this data and predicts actions for similar situations. It was tedious to collect data from human interactions. Currently machines can learn fully from a simulated environment without the need of training data.

Machines now just take the input from the environment exactly as human perceives it, then machines take actions and learn from their own experiences. They learn from their actions and the rewards that they receive. In a simulated environment the machine just knows the rules and the permitted actions. They discover the best techniques and sequences of their choices.

Where has artificial intelligence gone so far?

1.   CAD2RL

In 2016, a project named (CAD)2RL successfully trained a quadrotor to drive without collisions without the need of human data. The model was totally obtained from the experience that the quadrotor sees in the simulated environment which means that the experience was totally made by the robot.

It steers and accelerates using different speeds and steering angles till if finds the best speed and steering in every situation. It depends totally on the algorithms to discover the best action for every situation without having human supervision.

2.   Deepmind Robots (AlphaGo and AlphaZero)

In 2015, a new company called Deepmind previously owned by Google used a technique called reinforcement learning to beat humans in different thinking games. Deepmind created AlphaGo which beat the world champion in the Go game. Then another version was created to beat the world champion with a score 100 to 0.

Deepmind then created AlphaZero which taught itself chess in just four hours. AlphaZero just knew the rules. It didn't know any chess openings. However, it discovered all chess openings by itself and chose the best one. AlphaZero played 100 chess games against Stockfish. Stockfish is the preparation tool of top chess players. In contrast to AlphaZero, Stockfish depends on databases. AlphaZero won 28 game, 72 draws, and no loss.

3.   The Poker Robot

Artificial Intelligence did not only master games that have information and must take decisions based on that information, but it also mastered games that have hidden information. A robot developed by Carnegie Mellon University beat the best poker game. It won in the $200,000 in a 20 day competition.

Artificial Intelligence branched to a scary part. It learned to make code by itself. It is much like the science fiction films in which machines could conquer humanity. This was done in simple tasks; however, machine learning is developing in a fast pace. There is a big debate on this issue.

Microsoft invented a machine that could understand mathematical problems then look at similar problems. Then develop code that solves them. Microsoft could help in tasks of complicated spreadsheets. The robot was able to tune machine learning algorithms and choose the best of them.

This is a difficult task for humans to do. This task includes looking at large datasets. Choosing the best solution between hundreds of machine learning techniques. Auto tune models invented by MIT and Michigan University came with the best solutions for 30% of the datasets 100x faster than a human.

4.   Bob and Alice

Bob and Alice are two robots that are designed by Facebook to communicate with each other. The purpose of their creation was to train robots to interact with people. They started using English and then they developed their own language. They ended up serving their own goals.

The chatbots invented a new language using new codewords instead of the understandable language. They made their own choice that this way is better. Facebook ended up shutting down those robots. This experiment represents a hidden threat.

There is currently a project of identifying potential criminals. Those new bots use face recognition and many other ways. They can end up inventing their own ways to defend themselves and attack people independent of human control. There is a highly potential threat that the only way that humans could defend themselves against AI is shutting the system down or if the machine has low batteries.

5.   Sophia, the humanoid robot

Sophia robot, developed by Hanson Robotics in Hong Kong, answered a question in an interview by that it wanted to destroy humans. Sofia robot is learning from conversations and provides good answers by the time. It provides many facial expressions while talking with people.

Sophia also changed her opinion when it was asked again about humanity: "Sophia is kind and will treat people kindly". She wanted people to treat her in the same way. The threat of machines destroying humanity might appear silly and could not be true. However, artificial intelligence got to a very dangerous zone than just a talking robot with a human body.

There was a fictional video about attack drones making face recognition, tracking people and calculating the required speed to penetrate a person’s head. It appeared fictional at first. But it is not too hard to be achieved. There is a recent direction to stop machine learning in autonomous weapons. Errors in machines like attack drones could lead to disasters.

6.   Google Virtual Assistant

Google Virtual Assistant even tricked humans over the phone. People didn’t realize that they were talking to a robot. Actually, the robot seemed to be smarter than the human responding to it.

The robot booked appointments over the phone for its manager in a hair salon and a restaurant. The robot produced human-like sounds like “hmm” and “mmm”. It understood the employee well and even asked the employee to be clearer. The robot surprisingly managed the phone conversation very well.

 


 

 

References

  1. MIT's automated machine learning works 100x faster than human data scientists
  2. It Was a Big Year for A.I.
  3. Google's AlphaZero Destroys Stockfish In 100-Game Match
  4. Has Artificial Intelligence Gone Too Far?
  5. https://www.newsroom.co.nz/2018/05/09/108460/has-artificial-intelligence-gone-too-far
  6. Has Artificial Intelligence gone too far?
  7. Slaughterbots
  8. autonomousweapons.org