Post No. 22: Artificial Intelligence, Part Two

The last post on artificial intelligence described how easy it is to use Microsoft’s Bing chatbot.  It also looked at two amazing chatbot responses from queries it was given. This post briefly looks at differences between standard software in everyday computers and AI software.

AI is not new in the last year or so, although its popularity in the public is fairly recent. For example, AI has been the engine behind the posts and ads that you see in virtually every social media application.  It also has made possible speech recognition, language translators and a host of other automated speech and text related applications. So the question is, what makes AI different?

Traditional computing is based on software that is fixed and does not change without human intervention.  Software engineers have greatly improved traditional software development through a host of ingenious techniques, but it is not AI.  Artificial intelligence has the ability to learn and adapt to data presented. The output is generally a description, prediction or prescription for humans to consider. In many ways, artificial intelligence is a misnomer because it could more aptly be described as augmented human intelligence (AHI). Many industries can, and will, benefit from AI/AHI.

The parts of AI that make it qualitatively different can be simplified into two basic components:

1. Mad Amount of Data (MAD)

2. Ability to learn from the data (machine learning)

One example of using machine learning on mad amounts of data is the dramatic emergence of AI chatbots.  These chatbots were created by having AI software learn human language based on the mad amount of written human language on the internet. This is referred to as the Large Language Model.  It used a thing called “deep learning” to develop the chatbots.  More on this later in the post.

It takes tremendous computing power and some human intervention to pre-train an AI. To use a simplified example, suppose you want the AI to recognize all pictures of cats.  A human labels many pictures with cats and some without cats. The AI uses the data and some simple instructions, called algorithms, to identify any picture with a cat. If there are mistakes, the AI engineer can provide feedback on just a few examples, and now the AI has become very smart at identifying cats!  A simple example of a human created algorithm can be found in social media.  The algorithm might be to give a very high weighting to whatever a user clicks.  So if you click ads for dry roasted pecans, the AI eventually learns you are a pecan person!

A more complex example of machine learning can be found in self-driving cars. These cars have many high resolution cameras and other devices that help it correctly respond to the task of driving. But to do this, it must “learn” how to accurately detect and appropriately respond to objects in the outside world.  This is accomplished by what AI scientists call “deep learning”.  A main feature of deep learning is a method of processing information similar to neurons in animal brains, called “neural networks”.  These neural networks are a complex and “deep” layering of nodes connected to one another, like human neurons.  The AI software must be “taught” to recognize edges of objects (represented by a node) to eventually determine the relevance of the object to its purpose: avoiding collisions.  Many layers of interconnected nodes are needed to accomplish its purpose.  Sometimes human “supervision” is needed, but most learning is accomplished without human intervention.

A good way to understand the deeper layers of machine learning is to consider how a human might respond to an ambiguous object while driving a car.  Imagine it is a foggy day and you see something moving at the side of the road about fifty yards ahead of you. Immediately you give this information a very high weighting and slow down as you approach it.  You have learned from past experience that something ambiguous moving ahead of you can cause a collision.  As you get closer you see the movement is just a flag flapping in the wind.   You relax and continue on.  The AI accomplishes this by having learned to put various weightings on information from nodes at different layers. If the video is ambiguous, the AI gives it less weighting and relies more on past learnings, I.e. it gives these deeper nodes increased weighting (an ambiguous moving object may cause a collision).  The AI has learned that, once it gets more precision from the camera, to give this video information a higher weighting.  So as the car comes closer and its cameras more clearly capture the outlines of a flag, it gives this more precise information a higher weighting and moves past the flapping flag.

The preceding example is over-simplified. AI deep learning software is vastly more detailed and complex.  But the example outlines the high level theory of how deep learning works.  It might be similar to describing how cars run because the gas you put in them produce controlled explosions  that create movement that makes the wheels turn.  In a similar manner, cars with internal combustion engines are vastly more complex and detailed than this description.

I agree with the idea that what AI provides is an amplification of human abilities to improve human well-being. So I like the term, augmented human intelligence, AHI.  Following are some examples where AI/AHI can be of tremendous benefit.  These are just the tip of the iceberg. There are many current ones I have not mentioned, and many yet to be imagined.

- Medical imaging

- New drug development

- Medical insights from a mad amount of medical data.

- Legal document retrieval and analysis

- Major efficiencies on large industrial projects.

Hopefully these last two posts provide a better appreciation and understanding of AI.  Part three, and probably last, post on AI will explore AI’s downsides.  In particular it will examine the possibility of AI replacing us humans.  To do this we’ll take a deeper look at our uniqueness in the animal kingdom and whether this can be replicated by AI.  It should be interesting.