From Symbols To Signals: Getting Closer To Machine Intelligence

Until recently, most artificial-intelligence researchers accepted the view that thinking consists of the manipulation of discrete symbols, such as the written or spoken language. With this understanding, they achieved a degree of progress, notably in machine understanding and generation of natural-language communication, in symbolic mathematics programs, and in the automatic proving of theorems and assertions by machines. Groups at MIT, Carnegie-Meilon, and Stanford dramatically extended theorem-proving work by symbolically encoding large amounts of knowledge into “knowledge bases,” which led to the development of the first expert systems. Despite these successes, the symbolic approach has proven insufficient. It has not adequately modeled intelligent behavior such as scene analysis or comprehensive speech recognition. Morever, even the so-called successful applications could not achieve a desirable characteristic known as graceful degradation, which is the ability to know when the limits of knowledge are reached and thereby to give up or fall hack on “common sense” in an attempt to fill in gaps. For example, humans successfully cope with ungrammatical and incomplete utterances. Machines usually cannot. Common sense would seem to require far more symbols than the human brain can hold and process in real-time (half a second). Experimental evidence from cognitive psychology and neurophysiology indicates that the mind cannot possibly process a “program” of more than about a hundred steps in this time frame. What’s more, in order to respond as humans do, symbolic programs would have to execute thousands of steps. The emerging consensus is that some form of sub-symbolic parallel processing must take place to account for the first line of intelligent behavior that occurs within the half-second horizon. This has led to a resurgence of an earlier school of artificial intelligence now known as neural networks or connectionism. Connectionism takes a sub-symbolic perspective, more concerned with signals than with symbols. The basic unit of connectionist models are simple processing nodes loosely modeled on human neurons and connected in patterns inspired by various types of neurons in the brain. Each “neuron” in a machine is connected to a number of neighboring “neurons both for in put and output. If the input from a neighbor is great enough, two things may happen: the “neuron” receiving the inputs may broadcast an “energy pulse” to other neighboring “neurons," as human neurons do along their axons; and the “synaptic connection” with the neighboring, “neurons” may be modified, either becoming stronger or weaker (the way connectionists account for learning). Neural network systems, instead of manipulating symbols into a “solvable” expression, undergo energy oscillations that eventually are attracted to a steady state or a series of states. Although neural networks are built from rather simple individual unite, they offer the opportunity for highly complex systems to arise, especially since the functions involved are generally nonlinear and often stochastic. It is the nonlinear nature of these functions, allowing for small differences to accumulate and form self-organizing structures, that may be the key to both explaining human learning and emulating it in parallel machines. The commercial applications of AI could be greatly extended by having self-organizing expert systems apprentice themselves to human experts and learn through example to replace or supplement the laboriously formalized approach necessary in symbolic expert systems development. However, before this is possible, the gap between symbolic and sub-symbolic processing needs to be bridged. To borrow the phrase that orginated in classical electrical signal processing, there must be a “signal-to-symbol transform.” While symbols are insufficient to completely explain cognition, no serious attempt to explain cognition can ignore them either. The most popular architecture currently for signal-to-symbol transform. uses hierarchical levels of “collection” whereby different candidate patterns compete in the "settling" process of the neural network on one level and become themselves inputs to a more inclusive network on another level. At each point in the process, what is a signal on a lower level becomes part of a pattern on the level above and, in context, is recognized as a symbol. However, such symbols, unlike those proposed by the “physical symbol system" theorists, are not discrete, inert units waiting to be manipulated. Rather, they are active energy patterns in a society of neurons attempting to settle into a momentary steady state. Many complex systems of this type do not have a steady state, however, but only a non-repeating pattern of infinite feedback (a“strange attractor”). This is one reason, perhaps, why Socrates and his students found the definitions of important abstract concepts so impossible to write down, but rather found their meaning in the pathways of discussions they stimulated. Joe Bosurgi is a senior member of the advanced technology group at Symwztec Corp., a Cupertino, Calif, software development firm.

From Symbols To Signals: Getting Closer To Machine Intelligence

Interested in reading more?

Become a Member of