What is AI?

I hope you are enjoying articles in the #AIFiresideChat. As promised, here is the next blog in the series, helping you to understand “What is AI” in business terms.

In 1957, when researchers first conceived the word artificial intelligence (AI), they had no experience with calculators, personal computers, internet, electronic documents, spreadsheets, word processors, and many more. They conceived AI to be a technology that would mimic human behavior. As human behavior radically changed in the last 60+ years, so has the vision of how AI will impact the world. It has moved from the vision of a “HAL”-like robot that can do everything that humans can do, to having AI capabilities that accelerate and enhance specific tasks humans do routinely. This shift is the move from the concept of general AI to narrow AI.

While it seems that we are going backwards, the AI community has found that this increased focus has driven the success in targeted areas such as language translation, document understanding, intelligent interactions, image recognition, chess, AlphaGo, and self-driving cars.

A common question that we get is how we define AI. I have found that it is best to take a practical approach—AI allows computers to perform human tasks efficiently and accurately. The intelligence is the ability to learn, adapt to change, and derive insights. Specifically, we look for seeing human-like actions from the AI—for example, computing difficult equations or simulations are very challenging, but would we consider that intelligent? On the other hand, if computers start to learn from experimentation and start summarizing and organizing emails into various categories, is that a sign of intelligence?

To help further understand AI, we can think about the capabilities we use to act, and how we think. The following is our thinking around these capabilities using three tiers, tied to how they evolved over time, as AI progressed in its maturity.

Tier 1—Apply rules. Computers can use rules to automate business tasks and reduce manual work. In the 1980s, AI technologies began to develop rule-based approaches including strategies for how to search, select, and organize rules. They converted decisions into a set of conditions and actions, where the action takes place in the presence of associated conditions, for example recommending products based on past purchases. As another example, the basic task of preparing accounting statements, reviewing banking transactions for fraud, or applications for anti-money-laundering are common use cases for rules. These rules are very effective in solving concrete problems but have their own limitations. First, they needed an expert to maintain them and grow them. In addition, over time, they reach a breaking point where the rules would become too complex to change and make further improvements. Tier 1 was instrumental to large scale enterprise applications in Sales, Marketing, Billing, Accounting and Human Resource Management which now carry large rule-based repositories for transaction automation.

Tier 2—Learn. Most of the statistical learning advances were initiated in 1990s and matured into major commercial offerings in the last 10 years. Learning was a critical evolution in the development of AI. It allowed a computer program to use expert-generated examples to find patterns on its own. For example, it could learn how to identify people, dates, locations, and corporate events in company news. It could also learn from an expert how to recognize a complex pattern, such as a termination clause in a contract, which may be expressed in many ways by contract authors. This removes the need for having a human to manage a complex encyclopedia of rules, and also allows the system to learn complex behaviors that may be impossible to express as rules.

The ability for AI to work on unstructured data is important—today more than 95 percent of business information is either unstructured or semi-structured. This include emails, chats, documents, forms, checks, images, embedded tables, and audio and video recordings. Natural language processing enables computer programs to learn human ways of interpreting unstructured text. Machine and deep learning enable computer programs to recognize patterns, such as recognizing handwritten text or faces.

Tier 3—Reason.    The final tier in AI capabilities is the ability to reason and is likely to propel AI to many new practical applications over the next decade. While learning mimics human perception, reasoning emulates how we connect the dots and synthesize solutions. This leverages the patterns and relationships discovered by the AI in the learning process. For example, a knowledge graph can explore equipment behavior under different conditions to identify test and repair procedures for malfunctioning equipment. We use learning techniques to identify equipment symptoms and related alerts, and then use advanced reasoning techniques to choose test and repair strategies. This ability to reason has made AI systems popular as they have been used for equipment diagnostics, consultative selling, traffic rerouting, assigning work, assessing risk, or approving applications.

Today, the second tier, learning, is one of the most exciting areas in AI. Machine learning, a concept that was first introduced in analytics in the 1980s, has had an explosion of activity over the past two decades. This is the machine’s ability to learn from data. This capability comes in a number of different flavors:

Supervised learning provides a mechanism to develop a model by learning from an expert. For example, if you are asked to create an algorithm to recognize tables in a contract and classify them into different types of tables—payment schedule, deliverables, milestones, and so on—you can use supervised learning. An expert starts with a large number of table images and tags them into different types of tables. A computer program is trained using these tags to develop a model for identification and classification of tables. The training is done iteratively. In each iteration, the model classifies a different set of tables, the expert provides feedback, and the model learns from feedback to improve its performance. This is very similar to how we teach new team members in a consulting organization. The expert provides examples and then watches as the associate identifies and classifies tables, providing feedback on where they missed a table or misclassified it.

Unsupervised learning involves pattern matching without providing labels—the expert is no longer annotating the input data. The computer program separates the data based on statistical similarities and differences, and the results can be interpreted afterwards by an expert to provide meaningful labels. For example, we can apply unsupervised learning on the travel patterns of residents in a city or county. It will identify common mobility patterns by day of week and time of day and create clusters of people. An expert can review the resulting clusters and interpret them as “9–5 workers,” "commuters", or “Saturday morning golfers.” As cities and counties monitor spread of Coronavirus, unsupervised learning is helpful in identifying “social distancing” patterns by crowdsourcing mobility data from devices and identifying areas of concern and potential action.

These two approaches, supervised and unsupervised learning, provide a bulk of machine learning models. However, there are new, ingenious variations that help enhance the learning process and tackle different use cases that were previously thought to be difficult for machine learning.

Supervised learning approach often involve a large annotated training data set. However, these data sets are effort intensive and sometimes difficult to come by. Transfer learning comes to our rescue when training data on a subject area is scarce but is available plentifully in similar areas. Let us say you are looking for a way to read equipment numbers engraved on the equipment surface. While finding thousands of annotated equipment engraving examples is hard, we can find training data for similar annotated examples of text in various fonts and hand writings. As long as these examples cover similar letters and symbols, the learning can be transferred from one learning example to another. The existing learned model becomes a starting point for the new model, which learns incrementally to tackle the new use case. Many public and commercially available models support faster learning through transfer learning.

Reinforcement learning works well in situations where the objective is to optimize results that can be quantified. Many of you may be familiar with the ancient Chinese game, Go, an abstract strategy board game for two players, in which the aim is to surround more territory than the opponent.

As humans play Go, they get better. Can computers do the same? They can learn to play by practicing the game a couple of thousand or maybe a couple of million times. After all, they do not get fatigued. AlphaGo’s team introduced AlphaGo Zero, a version created without using any data from human games, and it competes better than any previous version.

This use of computing power to overcome issues of complexity is significant for problems that are massive in size and permutations. Imagine running a national marketing campaign program with many market segments and hundreds of campaigns. Reinforcement learning can be used to offload the need for heavy expert labeling and training, to the simpler problem of employing experts to specify the rewards and governing the results. The training effort is now shifted to large-scale computation for the computer to work through the “how” on its own. See the Forbes article below for many emerging use cases in this area.

Deep learning provides us an ability to develop a model using a series of hidden layers. This is similar to how perception plays a major role in human learning. As a child, we start interpreting faces, objects, and scenes we see in our daily life, we collect various speech utterances, and we learn to read symbols and characters—both handwritten and typed. However, how we do this is hidden and not something that we can explain. We intuitively understand how we read rapidly through a contract and find relevant passages, but we cannot express this as simple rules. When we read sentences, we can even skip many words while digesting the intent and meaning of a sentence. Deep learning provides machines the ability to mimic how humans use perception, by utilizing hidden layers that are trained over massive volumes of data and assist the learning task. Today’s deep learning systems in document understanding are approaching accuracies similar to human levels, and with better consistency. For example, a deep learning model reads 100-page documents in less than 60 seconds and retrieves all the relevant passages of interest for risk professionals. Such an assistant can significantly improve the performance of busy professionals, who can now devote time on analyzing and synthesizing instead of searching and retrieving.

Deep reinforcement learning combines the goal-seeking capability of reinforcement learning with deep learning’s ability to use large chunks of hidden layers to model perception. In many financial and operational fields, experts are able to find ways to play a game better, using heuristics and experimentation to try different strategies to test what works, and exploit what they have learned. This is true for an options trader doing currency trading, a campaign manager matching promotions to micro-segments, or a busy salesperson improving his or her chance to maximize sales through improved route planning. We can use deep reinforcement learning on large number of experiments, employing goal-seeking strategies, and thereby learn to optimize results. It works best in situations, such as stock trading and campaign optimization, where the results can be clearly stated in quantitative terms, but have a myriad of choices and actions that change the playing field.

This article provided you with a definition of AI using three tiers - rules, learning and reasoning, which collectively represent AI techniques and approaches. Tier 1 and tier 2 provided us a way to incorporate human decision-making and learning in an AI system. Tier 3 represents our ability to reason and thereby extend the value of AI by connecting dots, creating new work flows, harvesting collaboration across many AI systems and emulating human conversations. We will introduce you to some of the maturing reasoning applications and technologies later in this series.

Related Articles:

Here are the past articles from #AIFiresideChat.