Zero Genie

One of the challenges for those tracking the artificial intelligence industry is that, surprisingly, there’s no accepted, standard definition of what artificial intelligence really is. AI luminaries all have slightly different definitions of what AI is. Rodney Brooks says that “artificial intelligence doesn’t mean one thing… it’s a collection of practices and pieces that people put together”. Of course, that’s not particularly settling for companies that need to understand the breadth of what AI technologies are and how to apply them to their specific needs.

In general, most people would agree that the fundamental goals of AI are to enable machines to have cognition, perception, and decision-making capabilities that previously only humans or other intelligent creatures have. Max Tegmark simply defines AI as “intelligence that is not biological”. Simple enough but we don’t fully understand what biological intelligence itself means, and so trying to build it artificially is a challenge.

At the most abstract level, AI is machine behavior and functions that mimic the intelligence and behavior of humans. Specifically, this usually refers to what we come to think of as learning, problem solving, understanding and interacting with the real-world environment, and conversations and linguistic communication. However the specifics matter, especially when we’re trying to apply that intelligence to solve very specific problems businesses, organizations, and individuals have.

Saying AI but meaning something else

There are certainly a subset of those pursuing AI technologies with a goal of solving the ultimate problem: creating artificial general intelligence (AGI) that can handle any problem, situation, and thought process that a human can. AGI is certainly the goal for many in the AI research being done in academic and lab settings as it gets to the heart of answering the basic question of whether intelligence is something only biological entities can have. But the majority of those who are talking about AI in the market today are not talking about AGI or solving these fundamental questions of intelligence. Rather, they are looking at applying very specific subsets of AI to narrow problem areas. This is the classic Broad / Narrow (Strong / Weak) AI discussion.

Since no one has successfully built an AGI solution, it follows that all current AI solutions are narrow. While there certainly are a few narrow AI solutions that aim to solve broader questions of intelligence, the vast majority of narrow AI solutions are not trying to achieve anything greater than the specific problem the technology is being applied to. What we mean to say here is that we’re not doing narrow AI for the sake of solving a general AI problem, but rather narrow AI for the sake of narrow AI. It’s not going to get any broader for those particular organizations. In fact, it should be said that many enterprises don’t really care much about AGI, and the goal of AI for those organizations is not AGI.

If that’s the case, then it seems that the industry’s perception of what AI is and where it is heading differs from what many in research or academia think. What interests enterprises most about AI is not that it’s solving questions of general intelligence, but rather that there are specific things that humans have been doing in the organization that they would now like machines to do. The range of those tasks differs depending on the organization and the sort of problems they are trying to solve.  If this is the case, then why bother with an ill-defined term in which the original definition and goals are diverging rapidly from what is actually being put into practice?

What are cognitive technologies?

Perhaps a better term for narrow AI being applied for the sole sake of those narrow applications is cognitive technology. Rather than trying to build an artificial intelligence, enterprises are leveraging cognitive technologies to automate and enable a wide range of problem areas that require some aspect of cognition. Generally, you can group these aspects of cognition into three “P” categories, borrowed from the autonomous vehicles industry:

• Perceive – Understand the environment around you and input coming from sensors.
• Perception-related cognitive technologies include image and object recognition and classification (including facial recognition), natural language processing and generation, unstructured text and information processing, robotic sensor and IoT signal processing, and other forms of perceptual computing. Perception-focused capabilities is the area of AI research that got the biggest boost from the development of advanced neural network approaches, and Deep Learning in particular.
• Predict – Understand patterns to predict what will happen next and learn from different iterations to improve the overall performance of the system.
• Prediction-focused cognitive technologies utilize a range of machine learning, reinforcement learning, big data, and statistical approaches to process large volumes of information, identify patterns or anomalies, and suggest next steps and outcomes. Neural networks are helpful here, but so are other ways of doing machine learning as well as even simpler approaches such as knowledge graphs and statistical Bayesian models. Prediction-focused cognitive technologies span the range from big data analytics to complex, human-like decision modes.
• Plan – Use what was learned and perceived to make decisions and plan next steps.
• Planning-focused cognitive technologies include decision-making models and methods that try to mimic how humans make decisions. Early attempts include expert systems. More recent methods use a range of approaches that are used in situations such as cognitive-enabled cybersecurity or loan decisions. Planning-focused cognitive technologies is the area that can use greater AI-general research to improve as currently machines lack intuition, common sense, emotional IQ, and other factors that make humans much better at planning and decision-making.

From this perspective, it’s clear that while cognitive technologies are indeed a subset of Artificial Intelligence technologies, with the main difference being that AI can be applied both towards the goals of AGI as well as narrowly-focused AI applications. On the other-hand, using the term cognitive technology instead of AI is an acceptance of the fact that the technology being applied borrows from AI capabilities but doesn’t have ambitions of being anything other than technology applied to a narrow, specific task.

Surviving the next AI winter

The mood in the AI industry is noticeably shifting. Marketing hype, venture capital dollars, and government interest is all helping to push demand for AI skills and technology to its limits. We are still very far away from the end vision of AGI. Companies are quickly realizing the limits of AI technology and we risk industry backlash as enterprises push back on what is being overpromised and under delivered, just as we experienced in the first AI Winter. The big concern is that interest will cool too much and AI investment and research will again slow, leading to another AI Winter. However, perhaps the issue never has been with the term Artificial Intelligence. AI has always been a lofty goal upon which to set the sights of academic research and interest, much like building settlements on Mars or interstellar travel. However, just as the Space Race has resulted in technologies with broad adoption today, so too will the AI Quest result in cognitive technologies with broad adoption, even if we never achieve the goals of AGI.