Everyone is talking about "AI" today. Whether you are looking at Siri, Alexa or just the auto-correction features of your smartphone keyboard, we do not create artificial intelligence for general purposes. We create programs that can perform specific, limited tasks.
Computers can not "think"
Whenever a company states that there is a new "AI" function, it generally means that the company uses a machine to build a neural network. "Machine learning" is a technique by which a machine can learn how to better perform a given task.
We do not attack machine learning here! Machine learning is a fantastic technology with many powerful applications. But it's not general purpose artificial intelligence, and once you understand the limitations of machine learning, you can understand why our current AI technology is so limited.
The "artificial intelligence" of science-fiction dreams is a computer-aided or robot-based type of brain that thinks about things and understands them like humans. Such artificial intelligence would be artificial general intelligence (AGI), that is, it could think about different things and apply that intelligence to different domains. A related concept is "strong AI," a machine capable of experiencing human consciousness.
We do not have such AI yet. We are not near it. A computer unit like Siri, Alexa or Cortana does not understand and think like us humans. It really does not "understand" things.
The artificial intelligences we have are trained to perform a particular task very well, provided that the human can provide the data that will help him learn. They learn to do something but still do not understand it.
Computers do not understand
Gmail has a new Intelligent Response feature that suggests replies to emails. The Smart Reply feature identified "sent by my iPhone" as a common answer. It should also be suggested "I love you" in response to many different types of emails, including business emails.
That's because the computer does not understand what those answers mean. It has just been learned that many people send these phrases in emails. It is not known if you want to tell your boss "I love you" or not.
As another example, Google Photos assembled a collage of accidental carpet photos in one of our homes. This collage was then identified as the latest highlight in a Google Home Hub. Google Photos knew the photos were similar, but they did not understand how unimportant they were.
Machines often learn to play the system
Machine learning is all about assigning a task and making a computer decide the most efficient way to do it. Because they do not understand that, it's easy to learn how to solve a different problem than you wanted on the computer.
Here is a list of fun examples of creating "artificial intelligences" to play games and set goals. I've just learned to play the system. All of these examples are from this excellent table:
- "Creatures bred for speed become very large and produce high speed by falling."
- "Agent kills himself at the end of Level 1, to avoid that he loses at level 2. "
- " The agent stops the game indefinitely to avoid loss. "
- " In an artificial life simulation where survival requires energy but birth entails no energy costs were incurred for new children who could be eaten (or could be used as companions to produce more edible children). "
- " Since the AI were more likely to be "killed" when they lost a game, it was beneficial to be able to crash the game's genetic selection process. Therefore, several AIs developed ways to crash the game.
- "Neural networks have been developed to classify edible and poisonous fungi. They used the data presented in alternating order and did not really get to know any features of the input images. " Some of these solutions may be smart, but none of these neural networks have understood what they are doing. You have been assigned a goal and have found a way to achieve that goal. If the goal is to avoid losing in a computer game, pressing the pause key is the easiest and fastest solution they can find.
Machine Learning and Neural Networks
In machine learning, a computer is not programmed to perform a computer specific task. Instead, data is fed in and their performance is assessed in the task.
A basic example of machine learning is image recognition. Suppose we want to train a computer program to identify photos that contain a dog. We can give millions of pictures to a computer, some of which contain dogs and others do not. The pictures are labeled whether they contain a dog or not. The computer program "trains" itself to see what dogs look like based on that record.
The machine learning process trains a neural network, which is a multi-layered computer program that goes through each data entry, and each layer assigns different weights and probabilities before making a final decision. It's modeled how we think the brain could work, with different levels of neurons involved in thinking through a task. Deep learning generally refers to neural networks with many layers between input and output.
Knowing which photos in the dataset contain dogs and which do not, we can run the photos through the neural network and see if they give the correct answer. For example, if the network decides that a particular photo does not have a dog, there is a mechanism to inform the network that it was wrong to make some adjustments and try again. The computer recognizes better and better whether photos contain a dog.
This happens automatically. With the right software and a lot of structured data that the computer can train on, the computer can tune its neural network to recognize dogs in photos. We call that "AI".
But at the end of the day you do not have an intelligent computer program that understands what a dog is. You have a computer that has learned to decide if a dog is in a photo or not. That's still pretty impressive, but that's all it can.
And depending on the input, this neural network may not be as intelligent as it looks. For example, if there were no cat images in your dataset, the neural network may not detect a difference between cats and dogs, and mark all cats as dogs if you share it on people's real photos.
What Is Machine Learning Used For?
Machine learning is used for all types of tasks, including speech recognition. Language assistants like Google, Alexa, and Siri are so good at understanding human voices because of machine learning techniques that have trained them to understand human language. They have trained on a huge number of human speech samples and are getting better at understanding which sounds fit which words.
Self-driving cars use machine-learning techniques that train the computer to detect objects on the road and to learn how to respond correctly. Google Photos is packed with features like live albums that automatically identify people and animals in photos using machine learning.
DeepMind of Alphabet used Machine Learning to create AlphaGo, a computer program that plays the complex board game Go and beats the best people in the world. Machine learning has also been used to develop computers that master other games well, from chess to DOTA 2.
Machine learning is even used for Face ID on the latest iPhones. Your iPhone sets up a neural network that lets you see your face, and Apple has a special "neural engine" chip that handles all the numbers for these and other machine learning tasks.
Machine learning can be used for many different things, from identifying credit card fraud to personalized product recommendations on shopping websites.
However, the neural networks created by machine learning do not understand anything. These are beneficial programs that can fulfill the tight tasks for which they were trained, and that's it.
Photo credit: Phonlamai Photo / Shutterstock.com, Tatiana Shepeleva / Shutterstock.com, Diverse Photography / Shutterstock.com.