Understanding (Artificial) Intelligence

 This piece in the Atlantic from a few months ago is a wonderful profile of Douglas Hofstadter and a timely exposition of an issue at the core of the artificial intelligence enterprise today.

I read Doug Hofstadter's great book, Goedel, Escher, Bach (or GEB, as everyone calls it) in 1988 as a graduate student working in artificial intelligence - and, as with most people who read that book, it was a transformative experience. Without doubt, Hofstadter is one of the most profound thinkers of our time, even if he chooses to express himself in unconventional ways. This piece captures both the depth and tragedy of his work. It is the tragedy of the epicurean in a fast food world, of a philosopher among philistines. At a time when most people working in artificial intelligence have moved on to the "practical and possible" (i.e., where the money is), Hofstadter doggedly sticks with the "practically impossible", in the belief that his ideas and his approach will eventually recalibrate the calculus of possibility. The reference to Einstein at the end of the piece it truly telling.

My main concern, however, is the deeper point made in the Atlantic article: The degree to which the field of artificial intelligence (AI) has abandoned its original mission of replicating human intelligence and swerved towards more "practical" applications based on "Big Data". This point was raised vociferously by Fredrik deBoer in a recent piece, and much of this post is a response to his critique of the current state of AI.

deBoer begins with a simplistic dichotomy between what he terms the "cognitive" and the "probabilistic" models of intelligence. The former, studied by neuroscientists and psychologists - grouped together under the term "cognitive scientists" - was the original concern of AI, which sought to first understand and then replicate human intelligence. Instead, what dominates today is the latter approach which seeks to achieve practical capabilities such as machine translation, text analysis, recommendation, etc., through the application of statistics to large amounts of data without any attempt to "understand" the processes in cognitive terms. deBoer sees this as a retreat for AI from its original lofty goals to mere praxis driven, in his opinion, by the utter failure of cognitive science to elucidate how real intelligence works.

The high visibility of the statistics-based "machine learning" approach is real enough. To a large degree, it is a matter of what is possible and lucrative. In its formative decades, AI developed a lot of computational tools that were theoretically promising but could not be applied for the lack of computational power and sufficient data. Today, thanks to Moore's Law and the Internet, we lack for neither, and the same statistical analysis that seemed so impossible twenty years ago now powers search engines, recommendation systems and the occasional winner of Jeopardy. However, the critique of AI and its failures by deBoer and others is profoundly misplaced as I will try to argue below.

The main problem with the critique (and others like it) lies in its definition of "intelligence". Intelligence isn't something explicit that dwells in the brain and must be explained in terms of simpler primitives; it is an attribute that we assign to behavior that is sufficiently productive or complex. There is no "there" there other than this attribute, and the main issue in both understanding and replicating intelligence is to focus on the embodied behavior rather than on some disembodied Platonic essence called "intelligence". Once this is done, the dichotomy between the cognitive and probabilistic views disappears.

The idea that "true understanding" of intelligence must go beyond "mere statistics" is nothing more than crypto-dualism hiding behind statements of principle. Dualism is the notion that the "mind" has an essence beyond the material composition of the "body" - the two comprising a mind-body duality. This belief is, of course, virtually identical with the belief in a "soul", "spirit", "psyche", etc., and is profoundly rejected by modern science. Most of those who study mental phenomena today - neuroscientists, psychologists, philosophers of mind, and, yes, AI researchers - believe that the mind is a product of the physical body, much as life itself is. Just as the physical phenomena affecting the body - notably disease - that were once ascribed to divine providence are now understood exclusively in material terms, so are mental phenomena - perception, cognition, awareness, memory, thought - being moved ever so slowly away from immaterial explanations to material ones, with a corresponding increase in possibilities of intervention and artificial replication. Though the tools for doing so are still very primitive and a full understanding of how the mind emerges from the body is still a distant goal, a lot of progress has been made - deBoer and others notwithstanding. The anatomical connectivity within the brain and with the body has been studied for more than a century, and is now clearer than ever before. But there has also been a remarkable revolution in understanding the functional connectivity of the system and its underlying mechanisms. Reports on this are widespread in the scientific literature, but many of the most exciting results are summarized in a wonderfully readable new book by Stanislas Dehaene called Consciousness and the Brain (yes, he uses the "C-word" and lives!). Through such research, it has become increasingly clear that the mind - including whatever we may term intelligence - arises from the dynamics of the multi-scale complex network called the body; that ultimately, all percepts, thoughts, memories and choices correspond to ever-changing patterns of activity over hundreds of billions of cells - in the brain and the rest of the body - and that "understanding" these phenomena is a task for physics and chemistry rather than philosophy. The question, then, is this: What is the most appropriate quantitative framework in which to study these physical phenomena? And this is where AI finds itself.

AI, invented by computer scientists, lived long with the conceit that the mind was "just computation" - and failed miserably. This was not because the idea was fundamentally erroneous, but because "computation" was defined too narrowly. Brilliant people spent lifetimes attempting to write programs and encode rules underlying aspects of intelligence, believing that it was the algorithm that mattered rather than the physics that instantiated it. This turned out to be a mistake. Yes, intelligence is computation, but only in the broad sense that all informative physical interactions are computation - the kind of "computation" performed by muscles in the body, cells in the bloodstream, people in societies and bees in a hive. It is a computation where there is no distinction between hardware and software, between data and program; where results emerge from the flow of physical signals through physical structures in real-time rather than from abstract calculations; where the computation continually reconfigures the computer on which it is occurring (an idea central to GEB!) The plodding, sequential, careful step-by-step algorithms of classical AI stood no chance of capturing this maelstrom of profusion, but that does not mean that it cannot be captured! The fundamental insights that have powered the recent renaissance of artificial intelligence - and yes, there is one - are that:

• Mental capabilities from simple pattern recognition to intelligence arise through an evolving, growing and learning networked complex system (i.e., the embodied animal) interacting continually with its environment.

• These mental capabilities are meaningful only in the context of a specific body embedded in a specific environment, and not as general, disembodied abstractions.

In this view, the physical form of the body, along with all its capacity for change, endows it with the capability to support perception, cognition and intelligent autonomous behavior when it interacts with its environment. This happens, not through the instantiation of abstractions but through the continual flow of information (in material form - ions and molecules) through the cellular network of the body, which is continually reconfigured by this flow, thus shaping flows in the future. And while the flow of information is from the past to the future on the physical scale, mental phenomena are sufficiently slow that the linear causality of the "sense-think-act" sequence is replaced by a circular causality where perception, thought and action all influence each other in the course of any single mental event such as making a choice or recalling a memory. Today, brain imaging technologies allow us to see these flows in progress, albeit very imperfectly and in ways that are difficult to interpret. As the technologies improve, however, these functional networks of the brain and the body will become ever more visible.

 This "embodied" view of the mind has several important consequences. One of these is to revoke the idea of "intelligence" as a specific and special capability that resides in human minds. Rather, intelligence is just an attribute of animal bodies with nervous systems: The hunting behavior of the spider, the mating song of the bird and the solution of a crossword puzzle by a human are all examples of intelligence in action, differing not in their essence but only in the degree of their complexity, which reflects the differences in the complexity of the respective animals involved. And just as there is a continuum of complexity in animal forms, there is a corresponding continuum of complexity in intelligence. The quest for artificial intelligence is not to build artificial minds that can solve puzzles or write poetry, but to create artificial living systems that can run and fly, build nests, hunt prey, seek mates, form social structures, develop strategies, and, yes, eventually solve puzzles and write poetry. The first successes of AI will not be Supermind or Commander Data, but artificial flies and fish and rats, and thence to humans - as happened in the real world! And it will be done not just by building smarter computer programs but by building smarter bodies capable of learning ever more complex behavior just as an animal does in the course of development from infancy to adulthood. Artificial intelligence would then already have been achieved without anyone "understanding" it.

To decide whether progress is being made in artificial intelligence, the place to look is not in the theories of cognitive scientists but in the practice of neuroengineers.  Remarkable advances are occurring in using the information within the nervous system to control prosthetics and in replacing parts of the nervous system with artificial circuits. While those trapped in an older view of AI may fail to see it, these advances are the most concrete instances of artificial intelligence. The classical view is clouded by a distinction between "thought" and "action" - again, rooted in the dualism-that-will-not-die. But once it is recognized that perception, thought and action are all fundamentally manifestations of the same basic phenomenon - patterns of activity over complex biological networks - it becomes clear that there is no essential difference in the brain generating a pattern of nerve signals to control an artificial hand and the brain generating a pattern of nerve signals corresponding to a thought or a memory. The error lies in insisting that these processes have pre-specified forms that must fit our algorithms. The body demands no such thing - it just learns to generate the most fortuitous patterns of activity across its networks and lives (or dies) with the consequences. The wisdom encoded in its DNA by four billion years of evolution, in its structure by years of development from a single cell to a full-grown individual, and in its cellular networks by years of learning enables the body to be right more often than wrong, and more right the longer it survives. Statistics are just a way to capture this wisdom. The body does so without math; we do it with math. But there is no essential difference.

For critics such as deBoer, statistics and understanding are dichotomous. If and when a function of intelligence - say, facial recognition - is replicated through statistical methods, these critics declare that the computer does not "truly understand"; it is doing "mere statistics". What they fail to acknowledge is the possibility that this may be exactly what brains and bodies are doing too!

The statistical engines that power Google and Amazon may not fit our stereotypes of intelligence, but at an essential level, they are doing what animals do: Capturing the statistics of their environment and adjusting the statistics of their own bodies to thrive in that environment. The real task for those studying intelligence is to understand the physical mechanisms by which this occurs. This is where the action really is in the study of intelligence, and this is where is has been since seminal thinkers like  Donald Hebb and Horace Barlow first began to theorize about how the mind might arise from the body. And, far from being mired in failure, this enterprise is making progress every day.