Building a Conscious Machine

Image SourceTrying to program consciousness into computers has been an ongoing multi-decadal abysmal failure. The "top-down" approach of programming artificial intelligence into digital computing architectures has been bogged down by the huge differences between how the human brain works to create the mind, and how the human mind works to create human artifacts.

Basically, humans are stupid. Humans are stupid for believing that their conscious, rational minds can encompass the complexity of their 100 billion neuron brains without getting their feet wet and their hands dirty in the blood, gore, and sinew of low level, bottom-up, emergent phenomena. A few researchers have been working from the netherworld of thought, and are making progress.

As Dartmouth neuroscientist and Director of the Brain Engineering Lab Richard Granger puts it, “The history of top-down-only approaches is spectacular failure. We learned a ton, but mainly we learned these approaches don’t work.”

Gerald Edelman, a Nobel Prize-winning neuroscientist and Chairman of Neurobiology at Scripps Research Institute, first described the neurobotics approach back in 1978. In his “Theory of Neuronal Group Selection,” Edelman essentially argued that any individual’s nervous system employs a selection system similar to natural selection, though operating with a different mechanism. “It’s obvious that the brain is a huge population of individual neurons,” says UC Irvine neuroscientist Jeff Krichmar. “Neuronal Group Selection meant we could apply population models to neuroscience, we could examine things at a systems’ level.” This systems approach became the architectural blueprint for moving neurobotics forward.

....The robots in Jeff Krichmar’s lab don’t look like much. CARL-1, his latest model, is a squat, white trash can contraption with a couple of shopping cart wheels bolted to its side, a video camera wired to the lid, and a couple of bunny ears taped on for good measure. But open up that lid and you’ll find something remarkable — the beginnings of a truly biological nervous system. CARL-1 has thousands of neurons and millions of synapses that, he says, “are just about the edge of the amount of size and complexity found in real brains.” Not surprisingly, robots built this way — using the same operating principles as our nervous system — are called neurobots.

Krichmar emphasizes that these artificial nervous systems are based upon neurobiological principles rather than computer models of how intelligence works. The first of those principles, as he describes it, is: “The brain is embodied in the body and the body is embedded in the environment — so we build brains and then we put these brains in bodies and then we let these bodies loose in an environment to see what happens,” This has become something of a foundational principle — and the great and complex challenge — of neurobotics.

When you embed a brain in a body, you get behavior not often found in other robots. _hplus

Other attempts to build a brain from the bottom up include the Swiss Blue Brain project. Blue Brain is trying to build the cortical columns of a rat, then perhaps the entire cortex of a rat. From there, who knows?

Jeff Hawkins' Hierarchical Temporal Memory starts at a higher level than Blue Brain, but still grapples with the low level, essential messiness of the birthing of thought.

The late Francisco Varela, and Mark Johnson, and others have struggled with the concept of the embodied mind for decades, while their colleagues in cognitive science and artificial intelligence were beating themselves up attempting the top-down approach to intelligent machines. In the robotics field, Rodney Brooks was among the first to take a bottom-up approach to building robot brains.

Some of our most brilliant scientists and engineers have crashed and burned in the attempt to program minds from the top down. The problem is a conceptual one, but it often takes decades of failed attempts before even the most brilliant researcher understands the source of his failure. High intelligence is no protection from conceptual blindness. Sometimes it only makes it worse.

In theoretical biology, there is the concept of autopoieses -- self organising, self constructing phenomena. Nanotechnology is learning the idea from biology, in an attempt at a pragmatic skipping over some basic steps in nano-construction by borrowing ideas from biology. Gerald Edelman -- a Nobel Prize winner in immunology -- took his mastery of biological ideas to cognitive science, and began applying autopoieses to cognitive machines. It was a good idea, and progress is being made with it.

Whether humans will learn to "grow minds" intact -- as a whole -- or whether they will grow mental modules that can combine with each other in various ways to create multiple kinds of minds, the concept of autopoieses will be key to the creation of intelligent machines.

No doubt we will apply modifications and elaborations to these incubated minds, using top-down programming methods, but the core intelligence will have been evolved. Most people of "between levels" status will never understand that their brains and their minds work differently. They don't need to understand.

For next level humans, the concept will be elementary, simply a starting point as intuitively obvious as the hardness of stone or the wetness of liquid water.

Intelligent machines are a distinct possibility for the near term -- twenty years or so. Of course, once intelligent machines begin to evolve and combine ... and re-combine ... and re-combine ... who can say where the process ends? That is why more intelligent, wiser, and broader perspective humans are vital to the future -- and very soon.

That's why we can't afford bad government, bad media, bad academia, and bad child-raising any longer. Because the clock is ticking. Despite the Obama depression, despite the global jihad, despite the looming intolerant Chinese hegemony ... the clock is ticking.

There are a lot of things that need paying attention to. Who will be paying attention in 50 years?

Neurosciences Institute 5, Carnegie Mellon 0

At the invitation of the Defense Advanced Research Projects Agency, we incorporated a brain of the kind that we were just talking about into a Segway transporter. And we played a match of soccer against Carnegie Mellon University, which worked with an AI-based Segway. We won five games out of five. That’s because our device learned to pick up a ball and kick it back to a human colleague. It learned the colors of its teammates. It did not just execute algorithms. _Edelman

When you are making brains for robots, you need to follow the right pattern for what the robot is intended to do. If the robot is supposed to perform a human function, or play a human game, its brain should preferably be patterned after that of a human, to some extent. From an interview with Gerald Edelman of Scripps Research Institute and the Neurosciences Institute:

By proposing the possibility of artificial consciousness, are you comparing the human brain to a computer?
No. The world is unpredictable, and thus it is not an unambiguous algorithm on which computing is based. Your brain has to be creative about how it integrates the signals coming into it. And computers don’t do that. The human brain is capable of symbolic reference, not just syntax. Not just the ordering of things as you have in a computer, but also the meaning of things, if you will....

What exactly is a brain-based device?
It looks like maybe a robot, R2-D2 almost. But it isn’t a robot, because it’s not run by an artificial intelligence [AI] program of logic. It’s run by an artificial brain modeled on the vertebrate or mammalian brain. Where it differs from a real brain, aside from being simulated in a computer, is in the number of neurons. Compared with, let’s say, 30 billion neurons and a million billion connections in the human cortex alone, the most complex brain-based devices presently have less than a million neurons and maybe up to 10 million or so synapses, the space across which nerve impulses pass from one neuron to another.

Our brain-based device learned to pick up a ball and kick it back to a human colleague. It did not just execute algorithms....

Why is this kind of machine better than a robot controlled by traditional artificial intelligence software?
An artificial intelligence program is algorithmic: You write a series of instructions that are based on conditionals, and you anticipate what the problems might be. AI robot soccer players make mistakes because you can’t possibly anticipate every possible scenario on a field. Instead of writing algorithms, we have our BBDs play sample games and learn, just the way you train your dog to do tricks.

Artificial Intelligence enthusiasts almost always underestimate the importance of machine architecture when they fantasise about "human level machine intelligence." Machine intelligence acolytes often seem to feel that consciousness and cognition can be captured in an algorithm, and installed in a wide range of machine architectures. Unfortunately for that effort, intelligence is not algorithmic. Humans, as "intelligent beings", devise algorithms in order to help machines and humans accomplish goals more efficiently. But the algorithms are artificial constructs usually designed for specific tasks.

The same type of mistake is frequently made by "uploading enthusiasts," who think that human consciousness will sooner or later be uploaded into a more durable matrix than the "meat brain" it currently resides in. Needless to say, the problem they think they are discussing is not what they believe.

For an example of what Al Fin thinks is the most realistic "uploading" concept so far, read John Scalzi's "Old Man's War," a work of science fiction.

Interestingly, some British roboticists are beginning to devise parallel methods of machine learning and evolution, loosely modeled on biological evolution. And a company in Japan is offering to construct a "baby you" robot that looks like the picture of you that you send in with your order and $2215. Perhaps if the British and Japanese roboticists were to get together, we may soon be able to buy youthful robots that looked like us, that could grow to imitate us and learn to take one's place at dreary official functions?

Are you thinking about "Pod People", or "Kiln People", or other fictional scenarios you may have read or seen? Don't be too quick to dismiss the idea. Life is more fun when we inject just a bit of poetry, fantasy, and magic.

Special Bonus: A Wired article looking at another "brain architecture computing" project centered at IBM Almaden, that includes researchers from Stanford, Cornell, Columbia, and UCal Merced.

Second Nature--Neural Darwinism

Gerald Edelman is a distinguished scientist and Nobel laureate, accomplished in multiple scientific fields. His most recent book, Second Nature, summarizes and brings up to date much of his thinking and research on the natural selection of neural groupings in learning and cognition.

Neural Darwinism rests on three points:

1. Variation: On the microscopic level, every brain is connected differently. “Neurons that fire together, wire together.” Brain activity is constantly adjusting and changing these connections so that no two brains are identical. This point is in sharp contrast to the supporters of innate syntax like Stephen Pinker, who argue that all minds are the same.
2. Selection: Some of the various connections formed are preserved through a selection process while others are lost.
3. Feedback: The selection process is the result of a feedback (“reentry”) process that evaluates some brain connections favorably and some less favorably.

Source.

Here are some reviews of Edelman's recent book from a few well known scientists and thinkers:

"Until this provocative book, I thought that Gerald Edelman was merely one of our greatest and most original thinkers in neuroscience. But now having read such a remarkable disquisition on the relationship between brain physiology, consciousness and knowledge as he presents here, I have become certain of something about which I had previously only wondered: he is also one of our greatest philosophers."—Sherwin Nuland, Yale University

"Edelman's Second Nature offers the mature synthesis of his reflections on brain and mind. Somehow, it is both intellectually satisfying and wise."—Antonio Damasio, author of Descartes' Error and Looking for Spinoza

"A remarkable contribution to the philosophy of the mind, Edelman's Second Nature breaks new ground to an age-old problem by launching brain-based epistemology. Original, lucid, concise, succinct: easily the best in the field."—Apostolos P. Georgopoulos, Regents Professor, University of Minnesota

“Dr. Edelman has done something unique in this book. He deals both with the important epistemological issues and the mechanisms in the brain that give rise to them.”—Avrum Stroll, University of California, San Diego

“In the tradition of John von Neumann’s The Computer and the Brain and Erwin Schrödinger’s What Is Life? Gerald Edelman summarizes his seminal contributions to our understanding of the human brain and the human mind. The reader is drawn into a conversation with a master, who is at once witty and wise.”—Howard Gardner, author of Changing Minds

"It was William James's dream that physiology, psychology and philosophy be joined into a single discipline, and in Second Nature, the latest volume in Gerald M. Edelman's seminal series of books on Neural Darwinism, this dream of a brain-based epistemology is brought closer than ever to realization. For anyone who is interested in human consciousness, this is required reading. "—Oliver Sacks, author of The Man Who Mistook His Wife for a Hat

"Second Nature is well worth reading. It serves as a bridge between the traditionally separate camps of ‘hard’ science and the humanities."—Richard Restak, Wilson Quarterly

"[Edelman] reviews the latest research in brain-based approaches to consciousness, creativity, and mental illness."—ScienceNews.org

Source.

Edelman's Darwin VIII Discovers Neural Synchrony

Chris Chatham at Develintel Blog has added another post in his series on neural synchrony and oscillatory activity. The latest one deals with experiments in artificial intelligence from Gerald Edelman's lab, the complex neural net machine named Darwin VIII. Interestingly enough, Darwin VIII demonstrated synchronous activity during training, so long as the model's connections were left intact. The following is a bit technical:

The details of the implementation are as follows: the physical anatomy of Darwin VIII includes a CCD camera for vision, microphones for audition, infrared detectors for navigation, effectors for movement, and a 12 unit Beowulf cluster for the number crunching. Darwin's synthetic neuroanatomy consists of more than 53,000 units and over 1.5 million synaptic connections, with layers corresponding to V1, V2, V4, inferotemporal cortex (IT), superior colliculus, and ventral tegmental area (including a dopamine-like neuromodulatory system, based on an algorithm similar to temporal differences). For simplicity, the primary visual layer responds preferentially to green, red, horizontal, vertical and diagonal lines; all subsequent visual layers are bidirectionally connected and have increasingly large receptive fields (until IT, in which representation is non-topographic). The robot's orientation is guided by a topographic activity map in the superior colliculus layer, which also receives direct excitatory input from tones with specific amplitude and frequency picked up by the stereo microphones (this represents an a priori drive or bias for "target tones"). The dopamine system modulates synaptic efficacy between itself and IT, as well as between IT and superior colliculus, with effects that last several processing cycles. All areas contain both recurrent excitatory connections and lateral inhibition.

Neural activity was modeled via a mean firing rate model, with one small addition: a phase parameter "provides temporal specificity without incurring the computational costs associated with modeling of the spiking activity of individual neurons in real-time." All synaptic connections were modeled as phase-dependent, such that new phases are chosen at random unless the a unit's presynaptic input phases surpass a threshold, after which phase changes are first sent through a nonlinear "squashing" function and then scaled by a phase learning rate. This implementation causes postsynaptic phase to be influenced in the direction of the most active presynaptic units' phases. Synaptic efficacy is modified both with traditional firing-rate dependent credit/blame assignment, as well as with phase-dependent credit/blame assignment, in which units with tightly coupled phases are subject to potentiation, and those with uncoupled phases are subject to depression.

The experiment was divided into training and testing phases; during training, Darwin autonomously explored an environment consisting of one target item and three distractor items which share multiple attributes with the target item. For example, if a red diamond was the target, red squares and green diamonds would be distractor items. At the beginning of each training phase (which was repeated for three different Darwin "subjects"), all weights were randomized. Throughout the training phases, sounds were emitted from speakers which caused Darwin to orient towards the target. In the testing phase, these speakers are turned off and Darwin is allowed to explore its environment for another 15,000 cycles.

The authors measured Darwin's ability to locate the targets in its environment: each simulated subject was able to do so over 80% of the time. As the authors point out, " It should be noted that successful performance on this task is not trivial. Targets and distracters appeared in the visual field at many different scales and at many different positions as Darwin VIII explored its environment. Moreover, because of shared properties, targets cannot be reliably distinguished from distracters on the basis of color or shape alone."
At each timestep in the experiment, the researchers took a "snapshot" of the activity in every unit, and the weight of every connection. Results showed self-synchronization among neurons with recurrent connections within only 15 cycles; when reentrant connections were lesioned, no synchrony occurred within 10,000 cycles. Most importantly, multiple simultaneous synchronous firing patterns were observed within the active units of both IT, superior colliculus, and the value system (dopamine) layer, with both more synchrony and higher firing rates in circuits corresponding to targets or target features; in their own words, "the simultaneous viewing of two objects clearly evoked two distinct sets of circuits that were distributed throughout the simulated nervous system and distinguished by differences in the relative timing of their activity." In other words, multiple polyphase patterns of synchronous firing can self-organize inside a network with the proper architecture and environment.

You can find more information from Edelman's group here.

Thanks to Chris for alerting us to this remarkable finding. Neural net models apparently show emergent behaviour similar to that of actual neuron groupings in brains. Will the scientists find more similarity the more faithfully their models copy biological brains? Quite possibly. Then what? Then you might find more neuroscientists using neural net models to test their theories of brain function. When the information flows in both directions, the potential for important breakthroughs increases.

A Rising Star in Neuroscience

In a recent Science and Consciousness Review, Giulio Tononi is featured in an article by Henri Montandon. Tononi's full article, "An Information Integration Theory of Consciousness," which is described by Montandon, is also available.

Giulio Tononi may be the type of scientist who can give Jeff Hawkins a run for his money in the race to create a workable model of consciousness.

Gerald Edelman is well known to students of neuroscience and consciousness. Less well known is his frequent collaborator, Giulio Tononi.This link will take you to some reviews of "A Universe of Consciousness", by the two scientists.

I had intended to write a longer feature on Tononi two weeks ago, but got sidetracked. Thanks to mindhacks.com for reminding me. In the future I plan to compare some of the key concepts of Edelman, Tononi, Hawkins, and other consciousness researchers, in order to find key differences and similarities.