Defining artificial intelligence (AI) is a complex task, often subject to evolving perspectives. Definitions based on goals or tasks can change as technology advances. For instance, chess-playing systems were a focus of early AI research until IBM's Deep Blue defeated grandmaster Gary Kasparov in 1997, shifting the perception of chess from requiring intelligence to brute-force techniques.
On the other hand, definitions of AI that tend to focus on procedural or structural grounds often get bogged down in fundamentally unresolvable philosophical questions about mind, emergence and consciousness. These definitions do not further our understanding of how to construct intelligent systems or help us describe systems we have already made.
The Turing Test
The Turing test, often seen as a test of machine intelligence, was Alan Turing's way of sidestepping the intelligence question. It highlighted the semantic vagueness of intelligence and focused on what machines can do rather than how we label them.
“The original question ‘Can machines think?’ I believe to be too meaningless to deserve discussion. Nevertheless, I believe that at the end of the century, the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.” - Alan Turing
In the end, it is a matter of convention, not terribly different than debating whether we should refer to submarines as swimming or planes as flying. For Turing, what really mattered was the limits of what machines are capable of, not how we refer to those capabilities.
Measuring Human-Like Thinking
To that end, if you want to know if machines can think like humans, your best hope is to measure how well the machine can fool other people into thinking that it thinks like humans. Following Turing and the definition provided by the organizers of the first workshop on AI in 1956, we similarly hold that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”
To achieve human-like performance or behavior in any given task, AI should be capable of simulating it with a remarkable level of precision. The renowned Turing test was designed to assess this ability by evaluating how effectively a computer or machine could deceive an observer through unstructured conversation. Turing's original test even called for the machine to convincingly portray a female identity.
Alternative Tests for Human-Level Understanding
In recent years, significant advancements in machine learning techniques, coupled with the abundance of extensive training data, have enabled algorithms to engage in conversation with minimal understanding. Furthermore, seemingly insignificant tactics, such as deliberately incorporating random spelling mistakes and grammatical errors, contribute to algorithms becoming increasingly persuasive as virtual humans, despite lacking genuine intelligence.
Novel approaches to assessing human-level comprehension, such as the Winograd Schemas, propose querying a machine about its knowledge of the world, object uses, and affordances that are commonly understood by humans. For instance, if we were to ask the question, "Why didn't the trophy fit on the shelf? Because it was too big. What was too big?" any person would immediately ascertain that the trophy was the oversized element. Conversely, by making a simple substitution—"The trophy didn't fit on the shelf because it was too small. What was too small?"—we inquire about the inadequacy in size.
In this scenario, the answer unequivocally lies in the shelf. This test, with heightened precision, delves into the depths of machine knowledge pertaining to the world. Simple data mining alone cannot provide an answer. This definition necessitates that an AI has the capability to emulate any facet of human behavior, drawing a meaningful distinction from AI systems specifically designed to demonstrate intelligence for particular tasks.
Types of AI
Artificial General Intelligence (AGI)
Artificial General Intelligence (AGI), commonly known as General AI, is the concept most frequently discussed when referring to AI. It encompasses the systems that evoke futuristic notions of "robot overlords" ruling the world, capturing our collective imagination through literature and film.
Specific or Applied AI
Most of the research in the field focuses on specific or applied AI systems. These encompass a wide range of applications, from Google and Facebook's speech recognition and computer vision systems to the cybersecurity AI developed by our team at Vectra AI.
Applied systems typically leverage a diverse range of algorithms. Most algorithms are designed to learn and evolve over time, optimizing their performance as they gain access to new data. The ability to adapt and learn in response to new inputs defines the field of machine learning. However, it's important to note that not all AI systems require this capacity. Some AI systems can operate using algorithms that don't rely on learning, like Deep Blue's strategy for playing chess.
However, these occurrences are typically confined to well-defined environments and problem spaces. Indeed, expert systems, a pillar of classical AI (GOFAI), heavily rely on preprogrammed, rule-based knowledge instead of learning. It is believed that AGI, along with the majority of commonly applied AI tasks, necessitate some form of machine learning.
The Role of Machine Learning
The figure above shows the relationship between AI, machine learning, and deep learning. Deep learning is a specific form of machine learning, and while machine learning is assumed to be necessary for most advanced AI tasks, it is not on its own a necessary or defining feature of AI.
Machine learning is necessary to mimic the fundamental facets of human intelligence, rather than the intricacies. Take, for instance, the Logic Theorist AI program developed by Allen Newell and Herbert Simon in 1955. It accomplished the proof of 38 out of the initial 52 theorems in Principia Mathematica, all without any need for learning.
The Need for Learning in Complex Tasks
Far more difficult is the task of creating programs that recognize speech or find objects in images, despite being solved by humans with relative ease. This difficulty stems from the fact that although it is intuitively simple for humans, we cannot describe a simple set of rules that would pick out phonemes, letters and words from acoustical data. It’s the same reason why we can’t easily define the set of pixel features that distinguish one face from another.
The figure on the right, taken from Oliver Selfridge’s 1955 article, Pattern Recognition and Modern Computers—shows the same inputs and can lead to different outputs, depending on the context. Below, the H in THE and the A in CAT are identical pixel sets but their interpretation as an H or an A relies on the surrounding letters rather than the letters themselves.
For this reason, there has been more success when machines are allowed to learn how to solve problems rather than attempting to predefine what a solution looks like.