For the uninitiated: What makes AIML so transformative

Introduction

If you are a critical thinker still uninitiated into AIML (Artificial Intelligence, Machine Learning) technologies, you are justified in questioning the hype around these technologies and for wondering whether AIML is a fad or something more substantive. If this is a dilemma that bothers you, this article could help.

The quick answer to the above dilemma is that AIML is no fad. AIML is indeed in the class of a few technologies that are both general-purpose and highly transformative. No industry, no part of a company value chain, and no consumer will be left untouched by AIML. To think of the kind of impact that it is likely to have, think of the massive changes caused by other technologies that are in the same class as AIML, such as Electricity and the Internet. The changes due to this technology maybe even more profound and far-reaching.

In this article, I take a stab at answering the following questions:

· What makes AIML such transformative technology?

· How does it work?

· What are its limitations in its current form, and what does it mean for its transformative potential?

What makes AIML such transformative technology?

At its core, this technology enables machines to learn to perceive and to predict. It means that this technology could help machines to learn to perceive, say shapes, predict whether the form coming towards it is that of a cat or a leopard, and based on this prediction, decide whether to go ahead and stroke the cat or to do a runner.

Let us consider what I just said. Until recently, these cognitive tasks of learning, perceiving, and predicting were quintessentially human capabilities — capabilities that distinguished us from the machines. Well, that is no longer the case. The machines can now do these, not quite as well as humans can, but in their limited way, they can.

So, what are the implications of this? If machines can learn to do these cognitive tasks, then the learning and doing of these tasks can be automated. With physical tasks and the computational & logical tasks already automated (see Fig. 1), cognitive tasks were the last frontier in automation. Now there is nothing that humans can do that cannot potentially be automated; this has enormous implications for societies, economies, firms, and individuals, and this is what makes AIML highly transformative.

While the earlier paragraph sounds stark, and in some respects, the situation is indeed stark, there are some important caveats. At present, the ability of machines to learn and to perform these cognitive tasks is somewhat limited. In order to understand what these limitations are, let us first see how machines learn.

How does learning happen?

This section is a short primer on how machines learn to perform these cognitive tasks. But before we go to the machines, it will be helpful to understand how humans learn.

An influential theory of human learning posits that we learn in a perception-action cycle (see Fig. 2) — a process that keeps going on as a feedback loop throughout our lives.

· We perceive based on information coming to us via our senses, we predict based on this information, we decide based on our prediction and take action, and as a result of the action, there is an outcome.

· Back to where we started in the loop, we perceive this outcome, and this perception of the outcome serves as feedback. This feedback enhances our schemas, which in turn improves our perception and our predictions.

· The learning is this phenomenon of enhancing our schemas. Schema is best described as a “latticework of mental models”, conscious and sub-conscious, that encodes our past learning and experience, and that works also as a processor and enabler for our future learning. Depending on the feedback, the schemas could be enhanced in one or more of these three ways — the existing schema could be reinforced, or it could be extended, or a completely new schema could be built.

Hopefully, the explanation above will make it a bit easier to understand how the machines learn. Given below (see Fig. 3) is a diagrammatic representation of how Supervised Learning, a type of machine learning, happens.

Learning: The first step is learning. A model is trained based on labeled data, which we sometimes refer to as the training data. In the training data, each input (the feature set) in the data is paired with the correct output (the label). The idea is that the machines will figure out the right pattern for input to output mapping based on the training data. And then, it will use this learned pattern to label (hopefully correctly) the unlabeled input data.

Doing: Once the model is trained, it uses the input fields (the feature set) of the unlabeled input data to predict the label (output), and then the system will use this predicted label to take action based on which there is an outcome. This outcome can then be fed back into training to enhance the model.

Schema and Model: The schema in humans is equivalent to a Model in Supervised Learning, each getting updated continually based on a “learning” feedback loop. But apart from this equivalence, there are significant differences between the two. The schema is essentially a hypothesized entity based on what we observe about human learning, and even if it really exists, there is little understanding of what exactly its contents are and how they are held in our brain and consciousness.

Limitations of machine learning

While the way the machines learn broadly follows the principles of the way humans learn, so far, machine learning is not quite a match for human learning. There are three ways in which machine learning is known to be inferior to human learning. The machine learning:

· still requires a lot more data than human learning

· can be fragile if the future is too different from the past

· does not transfer well across different contexts

In contrast to machine intelligence, which is Narrow (context or domain-specific), human intelligence is General, i.e., it operates seamlessly across a variety of contexts. These findings further underline the fact that human learning and intelligence is a complex phenomenon about which our understanding is still somewhat limited, so our ability to replicate it in machines is only partially successful.

This technology is highly transformative even with its limitations

So, today’s AI is Narrow, i.e., context or domain-specific. While Artificial General Intelligence (AGI) is still far away, machines can be built to carry out cognitive tasks in a specific domain with a high degree of competence. Therefore, tasks requiring domain or context-specific cognitive skills that are currently performed by human beings can be automated. This Narrow AI has been successfully applied to Fraud Detection, Medical Diagnosis, Preventive Maintenance, Highway Driving, and many more uses cases across industry domains.

As the technology evolves and its predictive accuracy improves further, it will be applied to more and more use cases even more effectively. The Artificial Narrow Intelligence (as opposed to General Intelligence) can still be highly transformative.