Introduction to Artificial Intelligence (Part 1)

Our civilisation has been built through human intelligence. From agriculture, manufacturing, and communication to the device you are reading this article on, it has all been built by Homo Sapiens, which is Latin for wise man. What makes us humans special are our cognitive abilities and problem-solving capabilities. In fact, intelligence is such a fascinating topic; it has multiple definitions for what it is. If you ask me, I would say it comes in various types, such as verbal, emotional, logical and spatial, and how we can leverage these forms to think, act and solve problems.

If you are new to AI, you may think this is a relatively recent technology, but actually it dates back 70-75 years ago. We ought to start with Alan Turing’s paper in the 1950s, titled Computing Machinery and Intelligence, which essentially asks, ‘Can Machines Think?’ It is here where the Turing test was born. It involves a human and a computer that are asked a bunch of questions and a judge. If the judge cannot distinguish between the computer and human based on their responses, then the computer passes the test and is considered to be ‘intelligent’.

Now why is this case? Here we test the computer’s ability to understand and generate natural language, which is what gave rise to the field of NLP (Natural Language Processing). We test the machine’s ability to represent knowledge and store information and reason about this knowledge to draw conclusions. But really this is not enough. Additionally, we can perceive our environment based on what we see, and we can move and manipulate objects, such as holding a cup of tea, which I hope you have whilst reading this. So we need to incorporate computer vision and robotics for these respective capabilities, and this is what’s called the total Turing Test.

We will come back to what happened after that in a future article. Here let us try to dive deeper into what AI really is. Just like there have been multiple definitions of intelligence, there are various ways of thinking about AI.

Thinking Humanly

We need to study how humans think, and our understanding of the human mind is incomplete. However, we do have tools at our disposal, such as MRI, to analyse brain activity through scans. Additionally, we have this ability to introspect where we monitor and process thoughts as they come. You may think, should we build AI by imitating humans? After all, aeroplanes were not built to imitate birds; they were built through the study of aerodynamics, and so perhaps we can think in this manner too. However, we can most certainly use humans as a benchmark to evaluate the performance of AI systems.

Thinking Rationally

What does it mean to be rational? Essentially, it is doing the right thing. Although what constitutes being ‘right’ can be debatable. Just because we have separated it from human thinking does not mean humans are irrational, but we aren’t perfectly rational either. When we experience strong emotions like anger, stress, or sadness, our thinking can be clouded, and the outcome may be undesirable.

When we look at AI from this lens, we consider the laws of thought that lead to the study of logic and the syllogisms (structures) by which we define this logic.

Consider the statements: All humans are mortal, and Anuj is a human. You can easily conclude that Anuj is mortal. Notice, however, that for this to hold, we need these statements to be 100% true in the first place. What happens when we have uncertainty? Then it becomes difficult to formulate statements in this way. Logic like this requires statements of absolute certainty, which, of course, does not always hold in the real world. That is where probability fills this void. It is the study of uncertainty.

Acting Humanly

Agents are entities that act. It comes from the Latin verb agere, meaning to do. You’ll probably hear about AI agents in the media a lot, and really these refer to systems that can perform tasks autonomously to reach a goal. Autonomous because they plan, reason and decide the steps to take, which is different from automation, where we tend to define the steps and execution is automatic.

We talked about thinking humanly, but we don’t achieve things just by thinking. We need to perform actions. So you could think of intelligence as this blend of thinking and acting to achieve something. Although not everything requires thought. Do you have to think about pulling your hand away from a hot surface? No, it’s a reflex.

Acting Rationally

Here we strive to build rational agents that act to achieve the best outcome or the best expected outcome when we have uncertainty. We provide the objective to the agent, and it needs to learn how to best achieve this goal. This naturally leads to how we should define this objective in the first place. This can bring major challenges. A wrong objective can have negative outcomes, and even very intelligent systems can be problematic.

Consider a self-driving car. If we give a vague objective to reach our destination, it may do so without any regard for safety or road regulations and this creates a huge alignment problem between human preferences and what the self-driving car or AI in general does. Ultimately we want to provide the agent an objective plus any values that align with us.

I mentioned that very intelligent systems can be problematic. They may find loopholes and workarounds to solve a problem, but that may not be in alignment with what humans like. And again, this is why defining the objective is so crucial. The agent needs to meet our criteria rather than theirs.

Conclusion

So you can see that we can think about AI through multiple lenses, and hopefully you can see why this is a really interesting field and how we can understand intelligence this way. It combines many ideas from STEM (Science, Technology, Engineering and Mathematics) as well as economics, philosophy and psychology. In the next article we will look at the timeline of AI. How this field has evolved and the challenges faced in adoption and development. Stay tuned!