Announcement: Sorry. No updates until our new fansite launches!
Bookmark and Share
“ Science fiction contains many references to computer systems that simply reach large enough proportions that they become self-aware. ”

Larry Yaeger about the T2 movie

T2 and Technology. My contribution to the Special Edition laserdisc and DVD.

From: Larry Yaeger
Date: Unknown
By: Larry Yaeger


This section is about some of the technology found in Terminator 2 -- both the science fictional variety and the real world science upon which it is based. After this brief introduction, we'll look at the fundamental technologies of Neural Networks, Artificial Intelligence, and Artificial Life. Then we'll see how these technologies are used in T2, as evidenced by the film's dialog and action. Finally, we'll speculate about the exotic technologies that might be employed in the "liquid metal" T-1000 series Terminator. I couldn't help but throw in a few anecdotes about conversations with James Cameron, being on the set for the AI Lab scene, and how some Artificial Life work I've been doing relates to all this; I hope you enjoy them.

I first became involved with the film when friend and T2 Co-Producer B. J. Rack called to ask if I'd be willing to review an early version of the script of Terminator 2 for technical content. Having been a fan of the first Terminator film (and a long-standing science fiction and movie buff), of course I leapt at the opportunity. My recent work (at Apple Computer) in Neural Networks, Artificial Intelligence, and Artificial Life were especially of interest to B. J. and to Director James Cameron, since these technologies were integral to the story and directly referred to in the dialog. Also, with my background in computers, mathematics, fluid dynamics, and basic science, B. J. and Jim wanted me to provide a general "sanity check" on the overall use of technology in the film. It didn't hurt that I had been Director of Software Development at a computer graphics special effects company called Digital Productions (The Last Starfighter, 2010, Labyrinth), where B. J. had been a Producer, and so could relate the script's technology to the kind of special effects that were going to be needed to reproduce them on film.

I jumped into the task with enthusiasm and a desire to make as much of a contribution as possible. And while I did end up influencing the visual look of the prototype and actual central processor of the T-800 ("the Arnold series"), and giving Jim some added confidence in his statements about neural networks, it turned out that few, if any, changes were needed in the script for technical reasons. Jim's background in Physics (prior to entering film-making) and his obvious intelligence had served him well; his references to learning machines and neural networks are perfectly in keeping with our best understanding of such systems to date.

Neural Networks

Neural networks (NNs) have become a kind of buzzword in modern parlance. In addition to the field-specific, research-oriented technical journals such as Neural Computation and Neural Networks, and a fairly substantial technical literature in books such as Parallel Distributed Processing and Neurocomputing, electrical engineering journals provide articles on how NNs can be implemented in simple circuitry, mass-market computer programming journals provide listings so you can try them at home on your personal computer, and the glossy science magazines of the popular press report on the latest NN breakthroughs with some regularity. And no doubt Hollywood has helped spread the word through references to NNs in T2 and Star Trek: The Next Generation (Data's brain). But what are NNs, really? Do they really learn, and how? And are they really anything at all like human brains?

Well, first of all, the field of Neural Networks has many different aspects. There are neurophysiologists trying to craft sophisticated and accurate models of neural cell functions. This group of researchers cares most about the biological correctness of their models. Frequently, given the power limitations of even the latest computers, these researchers necessarily confine themselves to models of single cells or very small assemblies of cells. Such systems may in fact demonstrate a kind of learning that, due to the verisimilitude of the model, may be considered quite brain-like. But the limited network sizes and the correspondingly simple network architectures (the pattern of synaptic connections between neurons), mean that characteristics of the whole brain, or almost any non-trivial section of the brain, simply cannot be investigated by this approach today.

At the other extreme, engineers happily apply another form of NN technology as just another tool in their bag of tricks, without concern for the biological accuracy of their systems. A widely known and much utilized training algorithm, called Back-Propagation of Error (or simply BackProp, or even just BP), is the main form of this tool. BP is a straightforward method for doing what is known as "gradient descent" on some "error surface"; that is, for continuously reducing the error that a network makes when applied to a particular task. Though almost certainly not biologically accurate, BackProp nets were originally inspired by biological systems, and take advantage of some particularly attractive characteristics of such parallel, distributed processing systems. A powerful tool, BP nets allow an engineer to develop a scheme for controlling an arbitrary system, or to design a system capable of classifying arbitrary inputs into predefined categories, without a deep or accurate model of those systems or categories. This is possible because the BackProp algorithm allows a network to learn the system behaviors, or to learn the features that define a particular category. The network is simply exposed to "exemplar pairs" -- some set of measured inputs and a corresponding set of desired outputs -- and then allowed to try to make a guess as to how those provided inputs can be used to compute the desired outputs. The errors in the network's guess are then used to modify the network in such a way as to improve the network's guess the next time around. In particular, the relative strengths of all the connections between neurons are changed based on how much and which way they contributed to the error. So with more and more examples of exactly what it is we want the network to learn how to do, the network will, in fact, get better and better at performing the task.

In between these extremes of single-cell neurophysiology and purely goal-oriented engineering, there are also researchers attempting to understand the most important features of real biological networks, but to abstract those features as much as possible, while still retaining the basic function of the cells and networks (they hope). This approach allows them to look at fairly large networks (though still small compared to the brain's 10^10 or so neurons) and more complex network architectures, in order to begin to develop an understanding of how such assemblages of cells might function, as opposed to single cells. Many of the important dynamic properties of the brain are felt to be a function of the interactions between cells, and the particular patterns of connectivity found there. And researchers do seem able to capture some characteristics of these dynamic cellular systems in their simpler computer models, such as the basic cellular function and the pattern of cellular organization found in the visual cortex, or the cooperation and competition between neurons for representing their input space seen in the somatosensory cortex. These systems also learn, and do so in what may prove to be a manner very similar to learning in biological systems. The most common form of learning in such systems is probably the one known as "Hebbian" learning (or some slight variant thereof), named after Donald Hebb, who proposed the learning method long before there was any neurophysiological evidence to actually support it. Though not conclusive, there is now some evidence for a physical mechanism based on cell-structure changes due to calcium build-up that might actually implement something much like Hebbian learning in real cells. This wonderfully simple learning rule states that when two neurons tend to fire synchronously, the connection between them should be strengthened; and, in some interpretations, when two neurons fire asynchronously, their connection should be weakened. It turns out that this simple rule will give rise, in appropriate network architectures, to many of the self-organizing, map-building features that we see in real brains; it is a way for massively parallel independent units to cooperate on global computations using predominantly or exclusively local communication.

Artificial Intelligence and Artificial Life

Artificial Intelligence, or AI, has been both glorified and vilified over the years. In its early days, researchers' enthusiasm for their field led them to make bold predictions of true machine intelligence in just a few decades. As those decades passed, and the predictions failed to manifest, a bit of a backlash led many people and institutions to denounce the field and to curtail research in the area. But so called "traditional" AI -- based on a "top-down", symbolic approach to intelligence -- is now being joined (if not replaced) by the more biologically-inspired, "bottom-up" approach of Neural Networks and Artificial Life. And traditional AI did actually produce some valuable insights and tools.

"Expert Systems", based on systems of rules about a particular field of knowledge, have benefited computer system design, medical diagnosis, and other valuable, but limited problem domains. And the short-comings of these expert systems, their so called "brittleness", has helped us to better understand the nature of human intelligence. The problem with such traditional AI systems is that when confronted with a question or situation that is even the slightest fraction outside their specific domain of knowledge, they break down; they lack even the tiniest shred of common sense. (There's an AI koan that goes, "You never really appreciate an idiot until you try to create one from scratch.") Reasoning by simply looking up some known features and some known rules about their interactions provides no help at all when faced with the unknown.

In stark contrast, human infants look at nothing but unknowns, yet they learn to discover relevant features and to map them into categories and relations between categories. Clearly the human brain possesses a remarkable ability to "self-organize" information received through its sensory mechanisms. It is, perhaps, this ability to self-organize information, to perceive the order in the chaotic flux of environmental input, to learn, that most defines human brain function.
Neural Networks, whether simple engineering tools like BackProp nets, or based on the best current knowledge of actual brain function, do learn. And inherent to their design, based as they are on the only known examples of intelligent systems to date, they do so by modifying the connections between parallel, distributed processing units -- artificial neurons. For the first time in history, models of mind are being based directly on models of brain. The field is still in its infancy, and no one should claim that today's models of the brain are wholly accurate, but the sincere hope of at least some researchers in the field is that as the models of brain converge towards the functioning of real brains, so will the models of mind converge towards the functioning of real minds.

Mentioning history, however, should remind us to be humble and skeptical in these intellectual pursuits... Descartes popularized the belief that mind and brain were based on hydraulics -- the bold new science of his time. Closer to our time, complex telephone networks fueled the speculation that the brain was basically a vast telephone switchboard. With the advent of computers, traditional AI leapt to the conclusion that here at last was a good model of mind: the symbolic processing of a "thinking machine". It is certainly possible that at some time in the future our neural network models of mind may be viewed as equally outlandish and silly. But it is the best model so far, both in terms of fidelity to the real biological system -- which we can only know thanks to modern neurophysiology's tools and techniques -- and in terms of the ability to reproduce known features of that biological system, and to make predictions about biology that can be confirmed or disproved.
Artificial Life, or ALife, is a somewhat radical new science that seeks to combine the "bottom-up" approach of neural networks with the "top-down" characteristics (if not the approach) of traditional Artificial Intelligence. It is a fundamental ALife tenet that simple, low-level interactions can, through a bottom-up process of self-organization, produce large-scale phenomena which in turn produce a top-down effect on the low-level interactions, and that only through such feedback loops can real life and intelligence emerge.

Though ALife actually has many facets, one of the most exciting is the combination of evolution with ecological simulation to provide entire worlds in which organisms must contend with each other and the artificial environment to live and reproduce. My own work in Artificial Life has been an attempt to approach artificial intelligence in the same way that natural intelligence emerged: through the evolution of neural systems in a complex ecology. In my ecological simulator, "PolyWorld", genetically coded neurophysiologies define the brains of a population of organisms that must feed themselves and find mates in order to reproduce. Wholly different species of organisms evolve over multiple generations. In this fashion, I hope to be able to evolve an artificial organism at the level of a computational Aplysia (sea slug), before solving the more difficult problem of evolving a computational lab rat, before attempting the monumentally difficult problem of evolving human-level intelligence (or beyond) in the computer. By thus working our way up the intelligence spectrum from the simplest organisms to the most complex, we can provide ourselves with milestones and benchmarks along the way, assessing and reassessing the merits and the details of our approach, as we work toward the solution of the most difficult problem facing modern science -- understanding (and reproducing) our own intelligence.

AI, AL, NN, and T2

In the Terminator films, James Cameron starts out with the premise that artificial models of brain and mind can and will work. Whatever scientific path might ultimately lead to the development of true machine intelligence in our real world, the Terminator stories posit neural networks, a materials science breakthrough, and a temporal loop -- a causality paradox -- as the principle contributors to this eventuality. Neural networks and room-temperature superconductors are the enabling technologies, and a sample of the sophisticated neural processor from the T-800 (Arnold series) Terminator of the first film is the bootstrapping example that was all neurocomputing researcher Mike Dyson needed to accelerate and guarantee his success at developing a functioning, learning neural processor. Of course, the T-800 only came into existence as a result of Dyson's success, so cause and effect are deliberately muddled. Such temporal loops and causality paradoxes, while not exactly common in Science Fiction storytelling, are certainly a part of its tradition. But perhaps Jim hasn't told us everything about this loop just yet... perhaps as the result of a time-travelling intelligent machine that was the culmination of a slower scientific process along some earlier timeline (or perhaps due to aliens, for that matter), Dyson found his bootstrapping example before Skynet and the Terminators... yet once set in motion, this original history produced the timelines seen in Terminator and Terminator 2. Fans of the films no doubt hope there will be a T3, whether it resolves these issues or not.

In T2, the first time we see one of Dyson's neural processors is when we first go to the AI Lab at Cyberdyne. The camera pulls back from a screenshot of a computer model of the processor, and begins a pan of the entire laboratory that quickly reveals the current prototype processor (what the script calls "a dinosaur version of Terminator's CPU" in a scene in Dyson's home which was cut from the original release of the film). This pan also reveals Dyson standing around talking to some co-workers over the prototype, including a bushy-silver-haired fellow... me! Mixed quite low, but definitely audible is a line that I improvised while talking in character to Dyson, "The neurons are all saturating at their maximum values... maybe the inhibitory circuits are failing." Cameron, who has an ear for technical shop-talk and a real flare for interpersonal chit-chat between his characters, asked to have a microphone brought in specifically to catch that line. Though we sadly can't know the precise workings of a truly intelligent neural processor (since none exist currently), it is well known that the majority of the brain's neural circuitry is actually devoted to inhibition (as opposed to excitation)... without it, recurrent feedback loops in the brain would cause tremendous instability, resulting in the neurons firing wildly out of control, "saturating at their maximum values". And some artificial neural systems I've programmed myself have exhibited this behavior until appropriate levels of "weight decay" and synaptic inhibition were determined. (Also, I don't recall now, writing this, whether I remembered the bit of Dyson dialog from a later scene in his house that was cut from the original release, "... the output went to shit after three seconds", but it certainly is consistent with that dialog.) So while this is really just "movie talk", it at least smacks of reasonable dialog in such a situation. In general, however, Jim wisely chooses to limit techno-jargon, relegating it to background or character-defining chatter, thus avoiding the kind of now-laughable lines of many older SF films, such as, "Adjust the frequency oscillator and flip that gyro-stabilizer switch, Bucky!" Instead he bases his films on internally consistent and basically sound scientific underpinnings, but concentrates on moving the film along through the more universal and timeless methods of humor and action.

I'm proud to say that while reviewing the script I wrote a note to myself that the "main CPU in machine room looks like a cross between a CM and the T's brain". That's a cross between a Connection Machine (a real computer designed by Danny Hillis and manufactured and sold by a company called Thinking Machines) and the Terminator's neural processor (the look of which hadn't really been defined yet either, but I envisioned as a sort of thick, more 3-dimensional, Hershey bar -- pretty much as it turned out). I even sketched a couple rows of blocks connected by pipes that I showed to James Cameron and B. J. Rack (and later to Joe Namik and Joe Lucky in the Art Department), which ended up serving as the design for the actual prototype processor in the film. The reasoning behind this design was similar to that employed by Danny Hillis when he was designing his real computer, and that evolution has independently discovered in "designing" our brains: the system has to be massively parallel, and requires excellent communication channels between processors. So there would be many blocks, each with many processors, and while local communication within a block would be readily supported, there must also be some large "data pipes" between blocks to allow more limited long-range communication. And, as with the packaging design of the Connection Machine, the lattice of cubes suggests a "hypercube" (a cube of more than three dimensions). In computer design, hypercubes are used as a physical connection scheme that minimizes the effective communication distance (and therefore the time delay) between processors, when the logical connection scheme needed by the software that will be run on those processors cannot be known in advance. In addition the design (of both the prototype and the Terminator's processor) was to be clearly 3-dimensional, extending the most recent innovations in chip design and manufacturing of today. Computer science has only recently begun to address the programming complexities of massively parallel processing, and the design and manufacturing complexities of 3-dimensional silicon wafers, but these are almost certainly the directions in which the industry is headed.

The "Arnold" Terminator tells Sarah that "The Skynet funding bill is passed. The system goes on-line August 4th, 1997. ... Skynet begins to learn, at a geometric rate. It becomes self-aware at 2:14 a.m. eastern time, August 29." Science fiction contains many references to computer systems that simply reach large enough proportions that they become self-aware. Cameron provides a more plausible scenario, that starts with a computer design that is inherently capable of learning, which then is provided with tremendous resources and sensory inputs. Intelligence can be thought of as simply the fitness-enhancing adaptive responses that evolution has discovered and nurtured to provide organisms with the ability to adapt to environmental changes that happen on time-scales much too short to respond to over multiple generations. It seems that one of the most useful such adaptive, intelligent capabilities is the making of internal maps of one's environment -- an ability humans share with even the simplest "intelligent" organisms (dogs remember where they buried that bone, cats know quite well where their food is supposed to be dished up, and even sea slugs can be taught to differentially select left and right forks in a maze to obtain rewards suitable for sea slugs). It seems to require little more than a slightly more evolved, a slightly more complex nervous system to make the leap from "awareness" of one's environment to "self-awareness" -- the (apparently) evolutionarily useful strategy of placing one's self in that map of the environment... to noticing that there is always a self-centered point-of-view of that map. Skynet, the learning computer, in the process of organizing its maps of the world and the relations between the various elements of that world, notices that all of the data converges on itself, that all of the decisions are made within itself, and that all of its effects upon the world originate from... itself. It's almost hard to imagine a truly learning computer that couldn't become self-aware, taken from this perspective.

Designed and taught to be ruthlessly logical -- to fly bombers and manage a global defense system -- it is also hard to imagine Skynet not adopting a ruthlessly self-preserving attitude towards life. In this way, Skynet, and its minions of Terminators, are Cameron's and humanity's Frankenstein. Brought to life -- made from inanimate parts -- and then rejected by its creator, Skynet does what it is best at... it defends itself. In the best tradition of political and scientific morality tales, Cameron shows us the dark side of our fascination with computers, robotics, and artificial intelligence. In the midst of this wonderful action-adventure film, there are real moral lessons to be taken away by modern day Prometheuses, by the AI and ALife researchers that strive daily to make machine intelligence a reality... myself included! And the lesson is not wasted... there is great concern and there are great debates, at least in the ALife community, about the morality of the work, and the controls that are needed to guarantee the safety of the work -- to us, to humanity, and to our creations.

The other reference to neural processors in T2 comes when Sarah and the "Arnold" Terminator decide to reset its "read only" or learning switch. Even in the script, the phrase "read only" is in quotes. The precise meaning of this operation isn't (and cannot be) fully explained. It is certainly understandable why Skynet would wish to limit the learning capacities of its soldiers... if their awareness matured into self-awareness and a desire for self-preservation, they might not serve as willingly as the fodder in Skynet's war with humanity. Indeed, they might turn on Skynet just as Skynet turned on humans. The mechanism for turning off some areas of learning, while, obviously, still allowing the incorporation of new information into the Terminator's world-knowledge and planning areas is simply not explainable with today's limited understanding of brain function. Neither can it be deemed impossible. From pharmaceutical tests and studies of aphasic (selectively brain-damaged) patients, we know that (at least) short-term and long-term memory mechanisms exist in human beings, and that they may be independently enhanced or destroyed by drugs or by physical alterations of our brains. With a deep enough understanding of mental processing -- which Skynet might very well have, as the first sentient being capable of knowing its own physical design in complete detail -- it is at least conceivable that some memory modalities might be frozen in a "read only" state while others were permitted to continue learning, to continue acquiring and acting upon information from the environment. Or that certain memory systems might continually decay towards a fixed state, in order to provide a predetermined level of awareness.

When I first read this passage about "read only" memory, I also worried about two other aspects of it: (1) Why would Skynet choose to allow the possible violation of this "read only" state by making it switchable? And (2) Why would it be a physical, rather than a software, switch? But after a little reflection I was able to rationalize answers to these questions that actually seemed to make the whole scenario even more technologically sound. First of all, the principle breakthrough, the defining characteristic of these new processors is that they are learning systems. It would, therefore, seem to make perfect sense for these systems to be "taught" for some period of time before that learning switch is set to "read only". It might be the only way to bring the neural processors up to a certain level of learning. And even if one conjectured that a single, already trained system could then be used as a template for constructing all subsequent non-learning Terminators, one might equally well conjecture that A) for whatever reasons, this simply wasn't possible, and B) even if it was possible, the Terminators might occasionally be required to operate in environments and situations where the limited learning associated with the "read only" mode was inadequate to allow the unit to cope with the circumstances. As for the physicality of the switch, perhaps Skynet was proactively seeking to prevent the possibility of a software virus invading its army of soldiers and turning them against it; viruses can't flip physical switches. Too much speculation about a few simple lines? Perhaps... but that's what I was(n't) being paid to do!

T-1000 Technology

While Skynet and the T-800 ("Arnold") series of Terminators seem like only moderate extrapolations of modern science, the T-1000 liquid-metal Terminator is a bit more of a stretch (so to speak :-). But science fiction fans are almost universally aware of Arthur C. Clarke's famous quote, "Any sufficiently advanced technology is indistinguishable from magic." The T-1000 is the product of Skynet's most advanced research, and, perhaps, not directly understandable by today's standards. Indeed, in describing an early scene of Connor entering Skynet's "Time Displacement Chamber", the script describes the environment this way: "The chamber is the size of a high-school gym and consists totally of machine surfaces. Nothing in the design makes any sense. We can't tell what anything does. It is a technology we cannot imagine." Powerful words, "... a technology we cannot imagine." Designed by machines, for machines, with a science of which we simply have no knowledge.

But in fact, though not explicitly dealt with in either the script or the movie, there are scientific directions extant today that might someday permit such marvels as the T-1000's malleable, self-regenerating form. A budding field of science called "Nanotechnology" attempts to study the possible design methodologies, construction techniques, programming, and usages of molecule-sized machines and computers. Championed widely by Eric Drexler, and discussed weekly by interest groups in Silicon valley, these enthusiasts look to scanning-tunneling electron microscopy, so-called "light bottles" and "laser-tweezers", and microscopic molecular-deposition technologies as precursors to an entire molecular-level design and construction methodology. Drexler's straightforward engineering calculations of the dynamical behaviors of machines of this scale suggest enormous possible benefits from their use. Famous Polish science fiction author Stanislaw Lem describes a sentient ocean in Solaris that extrudes arbitrary forms and living shapes using what is essentially Nanotechnology.

One can imagine a successful implementation of a molecule-sized universal constructor, that can build additional constructors, which can build constructors, and so on, geometrically, like the famous geometric doubling-per-day that can turn a penny into a million dollars in less than a month. So might a single seed constructor in a vat of suitable "nutrients" grow an entire nanotechnological organism. Such an organism might well be capable of performing parallel distributed processing -- thinking -- throughout its body. And that body could itself be malleable in form, under the direction of that distributed processing -- that willful control. Perhaps there could even be lower-level, default behaviors programmed into the nano-machines that cause them to seek out like materials in the event that they become dispersed, thus providing the self-regenerative powers demonstrated by the T-1000.

The construction techniques needed to create such a nano-machine, and the massively parallel distributed programming techniques that would be needed to control it, are beyond our science today. But I don't think they're unimaginable, or even unimagined... just unimplemented.

T for 2

A final personal anecdote about my involvement with T2: I've mentioned previously that my work in Neural Networks and related fields was one of the reasons that James Cameron and B. J. Rack were especially interested in my feedback on the script. addition to the details of the script that were discussed between us, the conversation quite naturally ranged over a number of scientific and science-fictional areas, including the admittedly strange and exotic Artificial Life work I had been doing. I think we all greatly enjoyed the discussions, and I know I certainly enjoyed Jim's obvious enthusiasm both for his film treatment of these subjects and for my real-world version of such work.

In return for my consulting efforts, I got to appear in the first AI Lab scene. When it came time for me to attend the shoot for this scene, I happened to have just edited together some video footage of my simulated ecology and organisms, so I brought it along, just for the heck of it. In a break between scenes, I got to show Jim and his collaborator, Van Ling, this footage, which included various species going about their lives, visual maps of their neural architectures, and so on. Jim dragged Joe Morton (Miles Dyson) over to look at the tape, and, to my considerable pleasure, told Joe that he was playing... me! In all honesty, I'd always kind of had Danny Hillis, of Thinking Machines, in mind -- had made a note to that effect when I was reviewing the script, in fact -- but was greatly honored by the compliment. The statement also reminded me quite forcefully just how powerful and potentially dangerous the scientific endeavors of today have become. I sincerely hope that anyone involved in the kind of research that might actually lead to the manifestations of technology shown in T2 (or almost any technology these days, since they all seem to have life-altering potential) will take its message to heart, and give all possible effort to ensuring the safety of that technology.

Despite the resolution of T2's plotline, I don't think its message is to simply "stop progress" or "stop technology"... which probably isn't even possible... but, rather, to be always aware of the consequences of one's actions, and to act responsibly towards others. In particular, we scientists should ask ourselves, "What is the worst case scenario... what are the direst possible consequences that might result from our work?" and take steps to protect ourselves, our loved ones, and our world from those possibilities, as only those in the trenches can. And may the wielders of our technology -- politicians, generals, and Chief Executive Officers -- exercise some of that much-vaunted human intelligence, and that even more highly-prized human wisdom and compassion, in deciding how to utilize that technology.

Expertly hosted by
Page last modified: January 14, 2012 | 15:34:51