It's been a while since my last post on this particular topic. I got distracted by the state of my financies and trying to improve it (all the while forgetting that learning AI was motivated by learning skills to alleviate that). In this post, I'm going to pick up where I left off. (I've linked my previous post on the subject, in case anyone wants a refresher. If you haven't read it, I suggest you do that before reading this one.) That post ends off on categorising the types of human intelligence and determining the potential for AI to exhibit or operate within it. Currently, logical/mathematical  intelligence is the area with most potential, which isn't surprising given how computers are currently designed, built and operate.

Ways to Define Artificial Intelligence

As stated in my previous post, artificial intelligence has little to nothing to do with human intelligence. At best, AI can simulate some areas of human intelligence, but not truly replicate it. Since computers use algorithms to find goals, acquire data, process that data and achieve a result (without anything that can be termed understanding it), their goals and methods seldom align with human goals or methods of reaching them. Keeping that in mind, artificial intelligence can be categorised in four ways:

1. Acting Like a Human

When a computer acts in a sufficiently human-like manner (succeeds at making differentiation between the computer and a human impossible), it passes the Turing test. Due to the fact that popular media would have you believe that this is the be all and end all of AI, there are certain people who don't put much store in the Turing test.

The original Turing Test didn't factor in physical contact. The revised/updated Total Turing Test does include physical contact in the form of perceptual ability interrogation. (What that means in layman's terms is that the computer must incorporate/use both CV and robotics to pass the test.)

The aim is to achieve the goal, rather than to accurately/fully mimic human behaviour. (Both birds AKA spy drones and airplanes fly. They achieve the same goal, but by different methods.)

Do you fancy sex robots, anyone? Surely, I can't be the only one who does.

2. Thinking Like a Human

When a computer "thinks as a human", it performs tasks that require human intelligence/assistance as input (as opposed to rote performance of tasks). Driving a car is a case in point. In order to get a computer to think like a human requires a basic understanding of how we think. This is dealt with by cognitive modeling. It encompasses three techniques:

• Introspection: Monitoring one's own thought processes and documenting them
• Psychological Testing: Compiling a database of observed behaviour from a number of people in similar circumstances
• Brain Imaging: Directly monitoring brain activity through the use of Computerised Axial Tomography (CAT), Positron Emission Tomography (PET) and/or Magnetic Resonance Imaging (MRI) scans

Once a model has been created, it is possible to write an algorithm that simulates it. However, it is important to remember that there is a high amount of variability in human thought processes. This results in achieving accurate representation being highly difficult and the results of any code to do so being experimental in the best case and detrimental/disastrous in the worst one. This category of AI is often used for criminal/psychological profiling and related fields. Thus, accurately and reliably modelling human thought processes realistically is highly important.

3. Thinking Rationally

A person is considered to be rational when following a standard set of behaviours within a parameter range. A computer relies on the recorded standard behaviours in order to determine how to interact with the environment in a rational manner, based on the available data. The aim of using this approach is logical problem solving when it is possible to do so. This allows for creating a baseline technique, which can be adapted/modified to solve one or more actual problems. (The practice of solving actual real-world problems often differs from the theory).

4. Acting Rationally

By studying how a large number of humans act in certain circumstances and while under certain restraints makes it possible to determine the efficiency and effectiveness of certain techniques. For a computer to act rationally when interacting with an environment and various conditions and factors therein, it relies on accuracy in the recorded data, actions and algorithms it uses. As is true with rational thought, rational acts rely on a principle that may need to be adapted/adjusted in order to be applicable and/or useful in practice. That stated, it is possible to establish a baseline for a computer algorithm to achieve a goal by basing it on rational acts.

Human Processes Versus Rational Processes

As much as we might like to think otherwise, humans aren't always entirely rational (especially when influenced by emotion, instinct, intuition and other variable factors). Therefore, our processes differ from the rational processes of computers/AI. A process is considered to be rational if it always performs the right action and/or achieves the right goal based on the data/information currently available to it, given an ideal performance measure. (If rational processes are to go by the book, it must be assumed that the book is actually correct to begin with. We, on the other hand, sometimes throw the book out the window or even consider the data available to us.)

For example: The best/most rational way for everyone to drive is by always following the rules of the road. However, observing a typical major intersection for about half an hour will confirm that traffic isn't rational. If you follow the rules, you'll end up stuck somewhere because other drivers don't do so. (This is why rule-based navigation for self-driving cars almost always results in critical failure and/or the vehicle surrendering control to a human driver. Self-driving cars must act like humans, rather than rationally.)

Given the above, I am hoping that it will be possible for me to create and train an AI that can analyse, debug and test computer code to at least the same standard as I do. In doing so, the time I spend doing that aspect of my job can be freed up for actually writing new code.

Note: The categories used to categorise/define AI's behaviours and operations are useful for considerations of how to apply it to certain problems. However, keep in mind that these categories are rather arbitrary at best and have no definite boundaries. Certain schools of thought consider AI either "strong" (generalised and adaptable) or "weak" (performs specific tasks well, but struggles or fails with others). The problem with this is that a jack of all trades doesn't excel at any particular task, while weak AI can't reliably perform tasks without at least some supervision. Besides, having just two type classifications for AI isn't sufficient for a general understanding or decent classification.

There are four types of classification for AI, proposed by Arend Hintze:

• Reactive Machines: The machines in this category are the ones you see winning chess matches or Jeopardy! rounds. They have no experience or human-like memory on which their decisions are based. They rely purely on rule-based algorithms and computing power to analyse recorded past decisions and perform calculations (including probability). This is an example of a weak AI used for a specific purpose.
• Limited Memory: An autonomous robot (including a self-driving car) doesn't have the luxury of time required for making every decision (or a chain thereof) from scratch. Instead, these machines rely on a small amount of memory for recall and experiential knowledge of a number of varied situations. Reaction time is reduced by the machine being able to recognise the same or similar situation, thus providing more resources for making and recording new decisions (similar to how humans do). This is an example of strong AI as it currently exists.
• Theory of Mind: A machine capable of assessing not just its own goals and abilities, but also those of other machines and/or entities in a common environment has a degree of understanding that is currently feasible, but not yet commercially available. Until this level of AI is fully developed, self-driving cars will continue to fail and cede control. (They are required to not only get passengers from point A to B, but to also have the intuition to discern the potentially contradictory/conflicting goals of drivers and pedestrians in the environment and act accordingly.)
• Self-Awareness: So far, so good. According to films like those in the Terminator series, this is where the problems start. (Skynet was fine until it became self-aware, remember?) Fortunately for us, achieving this level of AI requires technological advances that aren't even remotely possible. Making machines aware of both themselves and others, as well as conscious and basing their intuition of others on experiential knowledge is currently beyond our capabilities. I hope the situation stays that way long after I'm dead.

Thumbnail image: AI thinking about itself, according to AI