Where does AI fit in today's business?

I once had a business client that had a data problem. He explained the problem to me and wanted to know if "machine learning" is the answer to the problem. Simply put, the problem was to complete columns in the dataset with missing values. In many of the rows, the dataset had lost so many critical values that there was not much hope to meaningful data recovery. I put together a solution that uses a semi-ensemble learning methodology to use existing data points and "estimate" the missing ones. There were excellent predictions in many of the rows but terrible results in some others. When we tested the method against data points that we had, the client was expecting to get estimations identical to original values or say a 99% accuracy. But, this was not always the case as the accuracy varied across the dataset. We tried many approaches but did not improve on the original one and concluded the project. The lesson learned from this experience was that high hopes in machine learning as a blackbox magic are toxic. So, why does this happen? What can "AI" really do?

What is AI in the first place?

Artificial intelligence can refer to machine intelligence and is also associated with the term "computational intelligence". Does artificial intelligence (AI) mean that intelligence is created artificially or an artifact has become intelligent? As of this writing, it's probably none of the two. Is machine intelligence (MI) a better term to use? That is, the intelligence we observe from a machine built by humans (or in future by other machines?). I prefer MI as a term to AI because it comes closer to reality. I will use the term AI to be closer to the hype, not the reality. What is the reality? In reality, AI started before all the fancy computing we see today. Let us define AI in its broader term.

AI refers to the computational behavior that are programed by humans and are exhibited by non-living objects (not necessarily artifacts) to accomplish specific tasks. This definition indicates that anything programmed by humans can exhibit a degree of intelligence of that human being or the person who taught that human being. As of the time of this writing, no machine has ever been programed to equalize the intelligence of a smart human scientist. That does not mean machines aren't capable of doing things that one single human cannot do. For example, a machine can compute the average of 100,000 real numbers in a fraction of a second while a human being cannot do that with the exact same algorithm. This process of transferring human intelligence to objects has started in ancient times. Mechanical arts were used to make things move and work according to predefined simple algorithms. For example, the design of natural cooling systems such as windcatchers exhibits some primitive algorithms that are executed by the windcatchers. The windcatcher is placed in a way that it automatically catches and then transfers the wind inside a building, aiming to cool it down. Later with the industrial revolution, intelligent artifacts were massively produced. For example, a watch is a mechanically intelligent system designed to automatically advance two watch hands in predefined steps of time. A washing machine comes closer to what we now perceive as AI, which is a mixture of electrical and mechanical system to automatically circulate clothes for a specific period. All these systems have been programmed by humans to do things automatically. So, they have some small degree of intelligence.

Why is Machine Learning Fascinating?

I may sound like someone who is trying to downplay the role of AI. AI is useful and has made our lives easier but I'm just trying to demystify the role of AI and our expectations from machine intelligence. When it comes to machine learning (ML), the objective is to move one step ahead of AI and make the machine automatically learn to do a specific time without a specific set of instructions (i.e., a fixed program of conditional statements). The objective of machine learning can be mildly achieved by using statistical analysis in the domain of the task to be learned so the machine can make decisions. These decisions are referred to as classification. A machine can be trained to identify images, sounds, or movements. All these tasks are possible based on a system of probability distributions. The machine can make statistical decisions based on the evidence it receives from its trainer, the human being. So, the human is not trying to program the machine into a fixed set of instructions. Instead, the human, for example, is trying to teach the machine to produce many models that capture the characteristics of human faces, animals, walls, cars, apples, and anything visual. This is a fascinating result because you can now see computers are able to distinguish objects, recognize voices, or speak fluently. But, is the result really fascinating? Not, if we go back to what intelligence is.

In its essence, the achieved machine intelligence, as of today, is extremely limited. There is no means of intrinsic detection of small surprising variations in the model. I have seen face recognition systems that cannot distinguish two siblings, although very easily distinguished by human beings. I have also seen that the system misclassifies two images that are not distinguishable by humans. All of these shortcomings are because the system is not really intelligent. It is just a model of parameters that are tuned to specific examples provided by the trainer. The intelligence gained is more akin to those possessed by trained animals.

What to Expect from the Future?

We cannot predict anything from the future. If we were able to do so, we would have changed the course of the history to our advantage. However, we can say that our current understanding of intelligence, which is a property of humans only, is so limited that the possibility of recreating the intelligence in an artifact that can by itself become intelligent enough to take control of our species and other machines is extremely unlikely in near future.