The popularity of artificial intelligence (AI) as some monster about to eat us all has gotten out-of-hand in my opinion. To help educate people then below I present some hopefully easy-to-understand diagrams explaining the state of AI today. Scroll below the images to see the explanation.
The hope is dispel the fear and near super-natural abilities of AI being foisted on the world today.
AI Diagrams
Step 1. The Problem Statement
Name the object displayed in an image when given an image showing only a single object.
In this example the image is an image of an apple. The solution is "apple" for the problem image.
Step 2. Ask The AI
The laptop represents a computer with an AI program. We are going to run a program where we select an image of a single object and then when the AI program runs the program will always respond with a name for the object.
Step 3. Guess
All AI programs guess. All of them. There exists a fancy word for guessing: probabilities. But these words mean the same thing. The AI program is going to guess a solution. The easiest way to conceptualize a computer guessing is to consider a random number generator. Further imagine the AI program has a file listing millions of words and names used in the English language on a separate line. The computer AI program simply uses a random number generator and then selects a word on a line based on the random number and then outputs the word on that line.
In our example the AI program rolled the dice and the word the random number word selected was "Bugs Bunny".
This AI program is extremely unhelpful. The program will almost assuredly be wrong and if on the off chance the word "apple" is correctly chosen then the output is not reproducible as it will choose randomly the next time. A random number generator is a terrible AI program.
The point here is to understand that computers can generate randomness. How computers create random numbers is a hard problem. However, creating a random number in and of itself is not considered AI.
Step 4. Learn
Machine learning is often associated with AI. Machine learning requires a feedback loop where incorrect guesses are fed back into the AI program such that the AI program does not make the same incorrect choice again.
In our example diagram the feedback loop is shown using an "Incorrect Basket".
Imagine you wanted to ask Siri on your iPhone "Find the nearest gas station", but instead said "Find the nearest filling station". Siri might reply, "The nearest Filling Station Night Club is 3 miles away on El Camino and Hillsdale." Then to correct your incorrect solution you say, "Find the nearest gas station." At this point Siri could store that "Filling State Night Club" as the incorrect answer for "filling station". Storing incorrect answers as incorrect is is called a feedback loop. This is also called machine learning.
Machine learning is a.) make a guess and then b.) store incorrect answers so as not to use them in the future.
To be sure, Siri could have also stored "gas station" as the correct solution for "filling station". However, in our example the AI was only told if the solution was correct or not. The solution of "Bugs Bunny" is not the correct answer for the apple image so we tell the AI the solution is incorrect. From our feedback loop then the AI will store the apple image along with the solution "Bugs Bunny" in the "Incorrect Basket".
In future apple image requests then all such apple images will be checked against the incorrect basket. If the new random word solution chosen by the random word AI program is in the incorrect basket then the AI program continue to guess until a solution is found that is not in the incorrect basket. One can imagine if one asked the AI program a million or so times to identify the apple image then eventually the correct solution will always be given because all of the incorrect solutions are in the incorrect basket. It will take forever too as all of the incorrect solutions have to be tried. Now if this seems like a slow way to get at the right answer then you understand why machine learning is so difficult and machine learning can lots of processing power, trial-and-error takes forever.
Step 5: Train
The next step in our journey of understanding AI journey is to train the AI program. Training in AI means to provide the AI with known solutions.
The diagram depicts training as putting two images with solutions into a "Correct Basket". We are pre-filling the solution basket with known solutions, thus the digram is labeled, "Pre-fill Correct Basket".
Sometimes in computer science we call this "hard coding" where we don't run some program algorithm to derive a solution, but instead we use a hard coded solution for the program to use.
And yes this is how all AI works. All AI machines that use machine learning use what is innocuously referred to as training data. I say innocuously because when the AI program comes up with a solution one is suppose to "ooh" and "aah" that some really complicated program derived the solution. Training data makes it sound better than "pre-configured solutions".
In this example we are pre-filling solutions for "apple" and "Bugs Bunny".
Step 6: Solve Known Solution
In this new AI program we are going to update our AI program to include a new step so as to check the "Correct Basket". The "Random Word" process still exists but is not displayed in this diagram.Given the apple image solution of "apple" has already been put in the correct basket then the AI program will correctly answer "apple".
Step 7: Solve Unknown Solution
In this step we put all the pieces of our previous steps together where we take advantage of both baskets, the "Incorrect" and "Correct" baskets. In addition we have a new process box called "Pattern Match Correct Basket" that replaces the "Random Word" process box. This process box has a single artificial intelligence rule:If the image contains the color red then guess "apple".
The "Bugs Bunny" image does not contain the color red. The first thing our AI program does is to check the "Correct" basket for the input image. However, the input image is unknown. The input image is not in the "Correct" basket.
Since a correct solution is not available then a pattern match process is used. An AI engine uses a bunch of if-then rules to characterize a solution. In this case there is one rule and that rule is if an image contains the color red then guess "apple".
Is the answer correct? Is the image an "apple"? Well, for sure it is not "Bugs Bunny". The determination of whether an "apple core" image is an apple is up to the human using the AI program. If the human tells the AI program the guess is correct then the solution goes in the "Correct" basket, otherwise the solution goes into the "Incorrect" basket.
Now You Understand AI
Now you understand the basis of AI and machine learning.Can AI intelligence beat human intelligence? Sure, easily because an AI engine can be pre-configured with far more solutions than a human can retain. An AI engine can also be configured with a much larger set of pattern matching characterizing rules than humans use. An AI engine can process information information much faster than humans.
Great, AI intelligence is better than human intelligence in this fashion. But this is no different than a computer has much more detail memory capacity than a human and a computer can store the entire Library of Congress on a hard drive where we cannot remember even a single book.
So what?
Now you understand AI hype
AI is a very popular topic in the media and entertainment industries.
The series of "Terminator" movies is all about AI going to war with humans.
Now that you understand the basis of modern AI then ask yourself does machine learning have the capability of free will or original thought to declare war on humans?
No. Guessing is not free will. Even if those guesses are so well informed due to a well fabricated set of pattern matching rules using an enormous mountain of known solutions the solution is just a guess informed by previous solutions and guesses.
Therefore AI is nothing to be feared.
Humans using AI to control systems where the risk of those guessed solutions are not sufficiently or accurately accounted for should be feared. But that is not about fearing AI, but fearing humans incorrectly using a tool like any other tool.
The fear of AI is not AI, but humans hyping AI as being something other than what is and then blaming AI so as to avoid human responsibility for failures and disasters.
Cheers!
-Mybrid
No comments:
Post a Comment