Can Machines Be Conscious?

So when you look at the dark screen, you rule out not just ”light” but countless other possibilities. You don't think of the stupefying number of possibilities, of course, but their mere existence corresponds to a huge amount of information.

Conscious experience consists of more than just differentiating among many states, however. Consider an idealized 1-megapixel digital camera. Even if each photodiode in the imager were just binary, the number of different patterns that imager could record is 2^{1 000 000}. Indeed, the camera could easily enter a different state for every frame from every movie that was or could ever be produced. It's a staggering amount of information. Yet the camera is obviously not conscious. Why not?

We think that the difference between you and the camera has to do with integrated information. The camera can indeed be in any one of an absurdly large number of different states. However, the 1-megapixel sensor chip isn't a single integrated system but rather a collection of one million individual, completely independent photodiodes, each with a repertoire of two states. And a million photodiodes are collectively no smarter than one photodiode.

By contrast, the repertoire of states available to you cannot be subdivided. You know this from experience: when you consciously see a certain image, you experience that image as an integrated whole. No matter how hard you try, you cannot divvy it up into smaller thumbprint images, and you cannot experience its colors independently of the shapes, or the left half of your field of view independently of the right half. Underlying this unity is a multitude of causal interactions among the relevant parts of your brain. And unlike chopping up the photodiodes in a camera sensor, disconnecting the elements of your brain that feed into consciousness would have profoundly detrimental effects.

To be conscious , then, you need to be a single integrated entity with a large repertoire of states. Let's take this one step further: your level of consciousness has to do with how much integrated information you can generate. That's why you have a higher level of consciousness than a tree frog or a supercomputer.

It is possible to work out a theoretical framework for gauging how effective different neural architectures would be at generating integrated information and therefore attaining a conscious state. This framework, the integrated information theory of consciousness, or IIT, is grounded in the mathematics of information and complexity theory and provides a specific measure of the amount of integrated information generated by any system comprising interacting parts. We call that measure Φ and express it in bits. The larger the value of Φ, the larger the entity's conscious repertoire. (For students of information theory, Φ is an intrinsic property of the system, and so it is different from the Shannon information that can be sent through a channel.)

IIT suggests a way of assessing consciousness in a machine--a Turing Test for consciousness, if you will. Other attempts at gauging machine consciousness, or at least intelligence, have fallen short. Carrying on an engaging conversation in natural language or playing strategy games were at various times thought to be uniquely human attributes. Any machine that had those capabilities would also have a human intellect, researchers once thought. But subsequent events proved them wrong--computer programs such as the chatterbot ALICE and the chess-playing supercomputer Deep Blue, which famously bested Garry Kasparov in 1997, demonstrated that machines can display human-level performance in narrow tasks. Yet none of those inventions displayed evidence of consciousness.

Scientists have also proposed that displaying emotion, self-recognition, or purposeful behavior are suitable criteria for machine consciousness. However, as we mentioned earlier, there are people who are clearly conscious but do not exhibit those traits.

What, then, would be a better test for machine consciousness? According to IIT, consciousness implies the availability of a large repertoire of states belonging to a single integrated system. To be useful, those internal states should also be highly informative about the world.

One test would be to ask the machine to describe a scene in a way that efficiently differentiates the scene's key features from the immense range of other possible scenes. Humans are fantastically good at this: presented with a photo, a painting, or a frame from a movie, a normal adult can describe what's going on, no matter how bizarre or novel the image is.

Consider the following response to a particular image: ”It's a robbery--there's a man holding a gun and pointing it at another man, maybe a store clerk.” Asked to elaborate, the person could go on to say that it's probably in a liquor store, given the bottles on the shelves, and that it may be in the United States, given the English-language newspaper and signs. Note that the exercise here is not to spot as many details as one can but to discriminate the scene, as a whole, from countless others.

So this is how we can test for machine consciousness: show it a picture and ask it for a concise description [see photos, ”A Better Turing Test”]. The machine should be able to extract the gist of the image (it's a liquor store) and what's happening (it's a robbery). The machine should also be able to describe which objects are in the picture and which are not (where's the getaway car?), as well as the spatial relationships among the objects (the robber is holding a gun) and the causal relationships (the other man is holding up his hands because the bad guy is pointing a gun at him).

The machine would have to do as well as any of us to be considered as conscious as we humans are--so that a human judge could not tell the difference--and not only for the robbery scene but for any and all other scenes presented to it.

No machine or program comes close to pulling off such a feat today. In fact, image understanding remains one of the great unsolved problems of artificial intelligence. Machine-vision algorithms do a reasonable job of recognizing ZIP codes on envelopes or signatures on checks and at picking out pedestrians in street scenes. But deviate slightly from these well-constrained tasks and the algorithms fail utterly.

Very soon, computer scientists will no doubt create a program that can automatically label thousands of common objects in an image--a person, a building, a gun. But that software will still be far from conscious. Unless the program is explicitly written to conclude that the combination of man, gun, building, and terrified customer implies ”robbery,” the program won't realize that something dangerous is going on. And even if it were so written, it might sound a false alarm if a 5â¿¿year-old boy walked into view holding a toy pistol. A sufficiently conscious machine would not make such a mistake.

What is the best way to build a conscious machine? Two complementary strategies come to mind: either copying the mammalian brain or evolving a machine. Research groups worldwide are already pursuing both strategies, though not necessarily with the explicit goal of creating machine consciousness.

Though both of us work with detailed biophysical computer simulations of the cortex, we are not optimistic that modeling the brain will provide the insights needed to construct a conscious machine in the next few decades. Consider this sobering lesson: the roundworm Caenorhabditis elegans is a tiny creature whose brain has 302 nerve cells. Back in 1986, scientists used electron microscopy to painstakingly map its roughly 6000 chemical synapses and its complete wiring diagram. Yet more than two decades later, there is still no working model of how this minimal nervous system functions.

Now scale that up to a human brain with its 100 billion or so neurons and a couple hundred trillion synapses. Tracing all those synapses one by one is close to impossible, and it is not even clear whether it would be particularly useful, because the brain is astoundingly plastic, and the connection strengths of synapses are in constant flux. Simulating such a gigantic neural network model in the hope of seeing consciousness emerge, with millions of parameters whose values are only vaguely known, will not happen in the foreseeable future.

A more plausible alternative is to start with a suitably abstracted mammal-like architecture and evolve it into a conscious entity. Sony's robotic dog, Aibo, and its humanoid, Qrio, were rudimentary attempts; they operated under a large number of fixed but flexible rules. Those rules yielded some impressive, lifelike behavior--chasing balls, dancing, climbing stairs--but such robots have no chance of passing our consciousness test.

So let's try another tack. At MIT, computational neuroscientist Tomaso Poggio has shown that vision systems based on hierarchical, multilayered maps of neuronlike elements perform admirably at learning to categorize real-world images. In fact, they rival the performance of state-of-the-art machine-vision systems. Yet such systems are still very brittle. Move the test setup from cloudy New England to the brighter skies of Southern California and the system's performance suffers. To begin to approach human behavior, such systems must become vastly more robust; likewise, the range of what they can recognize must increase considerably to encompass essentially all possible scenes.

Contemplating how to build such a machine will inevitably shed light on scientists' understanding of our own consciousness. And just as we ourselves have evolved to experience and appreciate the infinite richness of the world, so too will we evolve constructs that share with us and other sentient animals the most ineffable, the most subjective of all features of life: consciousness itself.