Artificial Intelligence, or AI, is a hot topic. There are various articles and stories touting the remarkable opportunities available to companies that leverage Artificial Intelligence (AI). Pundits point out that AI may be the most disruptive technological innovation since the steam engine and has the potential to drastically change the way we work, interact, and even live. You’ve heard from your CMO that you need an AI strategy and you’ve been talking to a myriad of vendors telling you why their AI solution is right for you. Before that can happen, you first need a common understanding of what Artificial Intelligence (AI) is, what it isn’t, and how it can fit within your overall analytical strategy. Within the automated intelligence framework, AI is a critical part of the overall analytical strategy, but it is not an end result itself. The focus for all analytical initiatives (whether Artificial Intelligence (AI) or not) is to automate business processes through analytics — not specifically deploying AI techniques in and of themselves. This document presents a strategy for positioning Artificial Intelligence (AI) within that context.
The operation of analytics and the processes they drive has always been the implicit goal of why we use analytical technologies in the first place. Moving forward, whether it is a sophisticated algorithm used to recommend offers or ad-hoc analysis to determine what campaigns work best and why, these all need to improve some business process. In other words, analytical activity does not stop at gaining insights or finding something interesting or generating a model score. Instead, it should be used to initiate or change some actionable phenomenon. The justification for this is based on three current trends that exist today, each of which continues to evolve. In this document we will delve into the field of Artificial Intelligence (AI) by understanding how it has evolved, how it is being leveraged today, and how you should think about it as you build an analytically driven solution within your own organization.
Recent Enhancements in AI
Although many of the current advancements are still overstated, what has gotten us closer to the initial vision of AI is the relatively recent convergence of storage technology, processing capability, and revised analytical methods. This evolving capability continues to find new ways to take advantage of new data types as well as the sheer volume of data. Although today’s Artificial Intelligence (AI) still doesn’t replace the need for people in most applications, algorithms can do things that were clumsy or unimaginable just a few years ago. For example, machines can now recognize handwritten characters as accurately as a human after seeing a single example. This is something that wasn’t possible until quite recently and with the emphasis on enabling technologies (including aggressive investments in Artificial Intelligence (AI) technology by countries like China) it’s almost assured that these capabilities will continue to evolve. This is why it’s considered one of the three major trends driving the analytics revolution.
AI uses a combination of human characteristics and sensory interaction to simulate human thinking and has evolved over many decades. Although Artificial Intelligence (AI) is a very broad and complex concept with a rich and varied history, many advances have reflected those of computational power. What was once impossible due to computational limitations has suddenly become possible. Current technologies have demonstrated many practical applications for these advances, but understanding these applications requires some basic understanding of what AI is and how it has evolved.
The Development of Artificial Intelligence (AI)
In the 1950s, many individuals envisioned developing computers that were capable of thinking like people, and they called this field of study “artificial intelligence.” Early computers were certainly limited by today’s standards and the advancement of any artificial “thinking” was similarly limited. For the most part, practitioners of the Artificial Intelligence (AI) concept used an ever-growing set of rule-based algorithms to create an expert system for one action or another. There were certainly benefits but learning was rudimentary. After several decades of trying to adapt machines to emulate the thinking processes of a human expert, computers still were unable to learn due to limitations. This led to the next phase of Artificial Intelligence (AI), where machines were taught to learn.
As a subset of Artificial Intelligence (AI), machine learning is not focused on teaching a machine to do something, but rather giving the machine the ability to learn through experience and improve its “knowledge” over time. Machine learning showed great promise as we began to manifest these capabilities in high profile ways such as IBM’s Deep Blue™ chess-playing success and later with Watson’s Jeopardy™ achievements. In both cases, the machine learned to compete at the highest level, but the ability to fully realize artificial thinking was still limited to very structured objectives.
For decades, the concept of neural networking existed but was not widely adopted for practical applications. Neural network modeling attempts to simulate the way the brain works by focusing on a specific outcome and then breaking it down into interconnected nodes and assigning weights to the interconnected relationships displayed between the nodes. The ability to do this repeatedly to improve output results is a powerful analytical method. However, early attempts at practical applications were marginalized due to the computational limitations of the time. In recent years, data scientists have taken advantage of advances in computational power to make these networks enormous and use them to process massive amounts of data. These networks are very effective when applied to the right use case. Thus began the branch of Artificial Intelligence (AI) and machine learning called “deep learning.”
Machine learning is a subset of Artificial Intelligence (AI) where algorithms are “trained” to get better autonomously from repeated scenarios. Next, deep learning is a subset of machine learning that takes a statistical approach to machine learning and applies long-standing concepts around neural network modeling to drive new and more complex applications for Artificial Intelligence (AI). These are not necessarily mutually exclusive or superseding techniques, but can support and build upon one another. The real questions for most people are – where are we with Artificial Intelligence (AI) today, and how much of it is hype versus reality?
How to Classify Artificial Intelligence (AI) Today
Many terms are used to classify artificial intelligence capabilities. Soft AI vs. hard AI, weak AI vs. strong AI, and narrow AI vs. artificial general intelligence (AGI) are contrasting categories that have been used with varying levels of popularity, with the latter probably representing the most commonly used designation today. Regardless of the terms, they all provide a clear distinction within Artificial Intelligence (AI) capabilities that provides for a clear delineation of when an Artificial Intelligence (AI) implementation moves from performing a set task towards having the ability to achieve outcomes using environmental context (i.e. truly emulates human intelligence).
Most pundits generally agree that True AGI is not currently available and there is no real consensus as to how close we are to this capability. Thus, the scope of the Automated Intelligence framework focuses exclusively on Narrow Artificial Intelligence (AI) concepts. Regardless of the terms commonly used to describe this discipline (soft, weak, narrow), nearly all Artificial Intelligence (AI) applications today are clearly in this category and many are very advanced in their capabilities. Whether it is something like facial recognition applications, online language translators, or even self-driving cars, these are all “narrow AI” applications because they still exhibit the traits identified above. Nevertheless, as is clear from the capabilities of these applications, “narrow” should not be synonymous with “simple” with regard to the deployment of artificial intelligence.
Narrowing Down Use Cases for “Narrow AI”
Unlike AGI or what some refer to as “Super AI,” narrow AI is still non-sentient in that it deals with computer models/applications/systems that are not self-aware of what they are designed to do, nor how they might address a problem based on unrelated past experiences. This is currently the level achieved by existing technologies, however, and can still result in some spectacular results. The following is how narrow AI use cases are classified within the automated intelligence framework.
• Machine learning models—individual algorithms meant to solve a specific use case independently
• AI-driven applications—applications that implement multiple AI algorithms and/or expert system rules to react to a specific set of environmental criteria
• Advanced interactive systems—multifaceted applications and models brought together to perform a complex service or interaction autonomously or in a new autonomous fashion.
There is a great disparity within this classification, from very simple to very complex, in terms of what each is designed to do. There is also a hierarchy in terms of how models are combined to develop Artificial Intelligence (AI)-driven applications. Next, models and applications can be brought together to develop advanced interactive systems, which may address many different use cases, but are still designed for a specific purpose and cannot evolve beyond that. Each of these is explained in more detail with specific examples in the following sections.
Machine Learning Models
As mentioned in the Evolution of AI section, pundits have commonly used the term “machine learning” to define the discipline of self-learning models using non-linear mathematical methods (such as neural networks) with a specific use case, often defined as “deep learning.” As the discipline continues to evolve, this binary definition will continue to evolve as some linear algorithms are very complex and some multidimensional algorithms are relatively simple; therefore, machine learning models can be better classified as falling into a complexity continuum. One available term to denote this range of sophistication is shallow learning and deep learning. In this category of narrow AI, we have a variety of general linear predictive models that fall into this continuum. There are also very complex and evolved use cases like natural language processing and face recognition that have an individual focus but are self-learning in nature. The way they would typically be represented in this range is as follows.
Individual linear algorithms
Most machine learning (ML) deployments involve individual predictive models that support a variety of business use cases across a broad spectrum of complexity using linear-based algorithms. Sometimes these models serve as inputs for more complex predictive solutions or are combined to create complete AI-driven applications (see Artificial Intelligence (AI)-driven applications). There are generally four types of predictive modeling techniques (which include multiple algorithm types for each).
The algorithms for these categories vary, and when in practice you may use multiple approaches before determining the right one. Denoting these as “linear” algorithms suggests that this group, at least for this category of ML algorithms, is reduced to non-neural network type algorithms. Two popular examples of solutions that use these “deep” learning capabilities include Natural Language Processing (NLP) and facial recognition
Natural Language Processing
NLP is an evolved capability that uses a number of techniques to simulate the human interactive process. Decisions are made almost instantaneously, but what goes into decisions includes the ability to interpret the situation, tone, predefined circumstantial rules, as well as experience (provided data is available to support it). Using NLP, one can create added value and efficiency by incorporating one or more of the following techniques.
• Automated summarization – This is the process of creating a short summary of a long piece of text that captures the most relevant information. Think of the abstract or executive summary found at the beginning of research papers and long reports. This can be achieved by extracting key sentences and combining them into a concise paragraph, or by generating a basic summary from keywords and phrases.
• Natural Language Generation (NLG)—This combines data analysis and text generation to take data and turn it into a language that humans can understand. While it is used to create jokes and poetry, it is also used to produce news articles based on stock market events and weather reports based on meteorological data.
• Speech Processing—This is the specific technology that allows virtual assistants to translate verbal commands into different actions for the computer to execute. This technology allows the Amazon Echo to translate your request to listen to some dance music into a specific Pandora search, or Siri to turn your question about local hot spots into a Yelp search for dinner recommendations.
• Topic Segmentation and Information Retrieval—These refer to the process of dividing text into meaningful units and identifying meaningful pieces of information based on a search query, respectively. You leverage this technology whenever you execute a Google search. Taken together, these two techniques are also being used by a number of legal tech companies to create searchable databases of legal opinions, allowing lawyers to find relevant case law more efficiently without having to spend hours poring over briefs.
• Biomedical text mining—This is a subset of text mining that is used by biomedical researchers to gain insights from vast databases of specialized research. Some of its applications include identifying relationships between different proteins and genes, as well as aiding in the creation of new hypotheses.
• Sentiment analysis—This is regularly used by social analytics companies to put numbers behind the emotions expressed on social media or the web to generate actionable insights. Marketers use sentiment analysis to inform brand strategies while customer service and product departments can use it to identify bugs, product enhancements, and potential new features.
Many techniques can be applied to fit a solution and many of these use cases have evolved to use very complex multi-dimensional algorithms (e.g. deep learning techniques) that allow the algorithm to continue to improve.
Facial Recognition Models
Another very popular use case that performs a simple task for people but uses complex mathematics is our third group of self-learning models, facial recognition. Facial recognition is a method by which software can identify a person or thing based on a digital image. Today these techniques work quite well on platforms such as Facebook and devices such as Apple’s iPhone. The ways in which these techniques work is basically a four step process.
• Recognize – The algorithm must first look at an image and identify each person’s face in real time. This may sound easy to do, but this technology has only been available since 2016 and is now found on almost every camera, smartphone, and social media site. The typical method breaks down pixels into gradients and then groups them into smaller squares that serve as a simple representation of the face and compares them to known patterns for faces in general.
• Posing – This step deals with the issue that the same face will undoubtedly be positioned differently in different images, yet still be (apparently) the same face. A common way to address this challenge is to identify key aspects of any face (rather than all pixels), then reposition these to assign a standard position or pose based on these facial landmarks.
• Encoding – This step identifies the key components of the face that are most relevant in recognition and is used to classify these as objects. Rather than trying to compare every unknown face to every known face (too much processing involved) or defining ideal measurements up front (too restrictive), most applications will give the algorithm the task of defining the most ideal measurements and then applying them. In this example, if a person looked at the measurements they would see a series of numbers that would mean nothing to them. Yet, the numbers and measurements that lie behind the numbers are the key to enabling this algorithm.
• Assignment – In this step we take a new image of an unknown person that we have encoded, find a known match, and apply the latter’s label to the former. This is a relatively efficient and accurate process because we have broken down the images into their most important features.
Read Also:
- Challenges And Future Of Adoption Of Artificial Intelligence (AI) In Educational Sectors
- Potential Of Artificial Intelligence (AI) In Healthcare
- From Admission To Discharge, How Artificial Intelligence (AI) Can Optimize Patient Care
- Advantages And Disadvantages Of The Use Of Artificial Intelligence (AI) In Management
- Artificial Intelligence (AI) Applications In Medicine
Leave a Reply