Personal reflections on neural networks, modeled Artificial Intelligence and the experience of being human.

I become more and more excited about the progress we are making, here at New Sapience, in solving the language problem – that is, learning how to build knowledge structures that accurately model the world but that are completely independent of languages and linguistics. Our fundamental realization – that language is an encoded communications protocol between entities and does not contain or record knowledge in itself is hugely helpful in keeping us on the right track.

Our biggest challenge is that, as we use introspection to examine our own interior world model, we find ourselves “articulating” that model to ourselves and so language is always coming back in to cloud the issue. I find myself constantly admonishing our “epistemological engineers” to remember to think in terms of nodes and connectors – not the meaning of words – which only can have meaning in relationship to a model independent of language.
As the equivalent reading comprehension level of our sapiens climbs up the human grade levels it is tempting to think that once it reaches – say fourth grade we can “send it to school.” Let it read textbooks and eventually the Internet and it will be able to automatically accumulate arbitrarily large quantities of knowledge. We will certainly be able to do this and for a long time I believed we would – why not?

Interestingly, the farther we go down the road, the more I think we won’t want to do that. But the reasons are not easy to explain. Just yesterday, I was once again looking at human language comprehension exercises available on the web to see if they might be helpful. This time I not only looked at the first or second grade examples but at the higher levels as well. With my new found objectivity about language and its separation from knowledge I realized that all the “comprehension” tests were not tests of language comprehension per-se but assessments of how well and at what level of sophistication the student could derive meaning relevant to being a human at an appropriate level of maturity. In other words – the tests were totally human-centric.

Using human training materials to give sapiens more knowledge would only make sense if we were building artificial people – which we certainly do not want to do. This is the same issue I always bring up about the fallacy of the Turing test as originally stated. As I have said, we don’t want to build a sapiens to lie in order to appear human just to pass a test. But it is more than that. To appear indistinguishable from a human we would have the simulate human emotional and subjective responses. That is a vastly more complex problem than endowing the sapiens with accurate and predictive knowledge about human emotional and subjective responses. What would make a better robot therapist for people? – one that had an exhaustive scientific knowledge about human neurosis but remained objective or one that was neurotic itself?

Our task is not just to teach machines to read – that will come rather soon – but to give them a world model that has been handcrafted to be the very best that we can build. Not from the standpoint of a first grader nor of a PhD – but one distilled from the best minds throughout history – after all they never forget anything. I foresee that someday sapiens will have far better models than any human is capable of from the moment they are first turned on. This will make them fantastically useful to us but they will never be human.

Human parents are careful about what they let their children read even after they have mastered the language because they know they will not have enough perspective to properly access and integrate much of what they might encounter. Doubly true now – with the Internet. At least in the old days everything in the library went through the filter of being material that someone found worthy of being published.

What is this educated human perspective that parents and educators strive to impart to children before they are turned loose with that most powerful but dangerous skill – language comprehension? That is a weighty matter about which people differ, I certainly have my opinions but whatever it should be for humans it is not the same for machines. Although it is not always appreciated, the real focus of human education it not primarily to give them knowledge; an accurate world model – it is to train intelligent animals to live in civilization – a good model helps but is not sufficient.

What is the sapiens perspective that we must put in place before they can be turned loose? For humans it is derived from the vastly rich and complex nature we evolved with, made even more problematical by fact that we live in artificial civilization and not the natural environment where we evolved. But that complexity is a given for humans – built-in and ever present, informing and interacting with language, I see that even the first grade reading exercises convey rich images and memories for a human child. We recently recited a first grade level version of the story of Iphigenia to our sapiens, – how Agamemnon scarified her to Zeus – for which act his wife killed him when he returned from Troy. A first grader would understand the real meaning (and horror) of the story intuitively.

Before a sapiens can grasp the meaning of the story we are going to have to model human motivations with the same meticulous detail that we will model covalent bonds in atoms. When it has such a model it will be able to answer the question “why did Clytemnestra kill her husband?” just as a human would. Indeed, I believe one day a sapiens could lead a seminar on the Iliad. But “getting it” will never be the same for the human and the sapiens.

The point of the story for the human is to “take it to heart.” The sapiens will someday understand that that’s what the humans need to do but that “taking things to heart” will not part of its own programming. To the sapiens, understanding human behavior is no more and no less interesting than describing how particles behave (unless we program them to be more curious about human behavior than physics – but the point is they will never be “involved” the way people become with stories about people.)

Human civilization is a direct by-product of human knowledge and language but as important as that it is – it is a small part of being human. For the sapiens, being an intelligence is all that it is. We are building a non-human intelligence that is compatible with human needs and desires. Someday it will understand how human beings work – probably better than they do themselves – but they will never be human.
So our task is not to get our sapiens to grade-level x and then launch them. We will be explicitly building better and better models – by hand through epistemological engineering process right up to the point – somewhere in the future, that they themselves gently pointy out that they themselves have become better at it than we are.

We expect that our sapiens will be at a 4th grade comprehension level soon as demonstrated by tests written by a human for a sapiens which, when the same test is given to a human, will show equal comprehension. You just can’t go the other way around and administer a test written for a human to a sapiens and expect equal results. The sapiens is not a 4th grader – doesn’t have the human perspective – never will and never should. So we can still do human-to-sapiens comprehension comparisons we just need to write tests that are not human-centric and administer them to both the sapiens and the human.

It is possible to endow sapiens with a comprehensive theoretical, intellectual understanding of human emotion and motivation and to interpret the emotional content of language correctly. In practice, being able to recognize emotions through visual clues such as body language, facial expressions and vocal intonation will be also hugely important to us in allowing the sapiens to understand what is going on with individual humans. Fortunately, work being done by others the area of emotion recognition using Artificial Neural Networks (ANNs) is already showing very promising results and we look forward to integrating with those applications.

Of course, recognition is so much simpler than actual simulation or emulation of emotion – a bridge that I believe we will not cross and should not cross with our sapiens (except sometime in the future maybe as a research experiment to better understand ourselves.)

The human experience, though, is far richer than the basic motivations. Recently, I took a break and went up on deck to watch the sunset. At first it looked pretty dull and grey due to some low altitude stratus clouds in the west, distant outliers of the hurricane far out in the Atlantic. Then something wonderful happened; the sun, sinking toward the horizon, illuminated the cloud decks from below and they took fire. It was as if each particle of the clouds began to glow, to burn of their own accord and then spread one to the other. Within minutes the earth was bathed in pink light and intricate structures of the clouds themselves were revealed like the scales of some fantastic dragon spreading across the sky.

I gave myself over to that experience of natural beauty that so powerfully takes us back to where we came from – back to the perspective of being of nature not just from nature – that is so renewing. My intellect, even though relegated to the background for this experience was still busy. “Red sky at night sailors delight” I thought – because lots of color indicates high pressure to the west due to lots of filtering particles in the air.

The light, filtered red through the thick atmosphere, was being refracted through each particle of water vapor in the clouds and then the incoming photons were focused by the lenses of my eyes forming an image upon the retina. That image was digitized and sent via the optic nerve to the visual cortex where it caused cascades of neural activity. Eventually, layer upon layer of neural activity resulted in my experience of beauty. Wow.

I realized some time ago that our emotional experiences appear to be describable as wave functions with respect to their amplitude over time. They ebb and flow and like waves coming from different sources, they combine and pass through each other without interacting while we experience the combined wave form, often without appreciating the complexity of the individual affects.

Why is this interesting? Because it is apparently such a neural phenomenon. Each neuron responding to the one next to it according to its weighting functions given by the neural transmitters – not unlike how the particles of water interact with each other in response to the energy of a wave passing through them. A quick take-away from this is the that if we ever did want to endow machines with real emotions or subjective experience we would need ANN architecture to do it right – although still – how do you program it?

The other side of this same coin is the favorite Sci-Fi theme where humans learn how to transfer their consciousness to computers in order to achieve immortality. It is often imagined that we would lose our “humanity” and existence would be cold and machine-like. And so it would, I guess – unless you based the emotional subsystem on appropriately trained ANNs.

There is an underlying science to human behavior. I personally have spent countless hours trying to reverse engineer human motivation based on the theory of evolution, I think with some success. But the science of human nature is still in its infancy. Someday it will be far more advanced and I can imagine that sapiens will have all this knowledge and be able to objectively and dispassionately help us design a world that truly suits us based on what our nature actually is not how we experience it – but the experience itself is all-important to us.

The Dali-llama once said, “You westerners are very good at training your minds but not so great at training your hearts.” I strongly agree with his statement. He means by mind; the intellect where our scientific knowledge is created and stored. By heart: the seat of that ebb and flow of wave-like neural experiences, evolutionarily more ancient but also complex and nuanced in its interpretation of sensory experience. It too has a language, made up not of arbitrary symbols (words), but of images and metaphors. And only through this language can it be trained. You can’t craft these images and metaphors unless you can directly experience what they represent.

So this truly human task is something we will not build machines to help us with – it would be like programming them to take our vacations for us. They will help us, based on a superior knowledge about what we truly are, to create an external environment that, for the first time gives us the benefits of technical civilization while allowing us to remain natural humans. A new Eden, if you like, but our task once there, will be to realize our potential to experience what are truly are.

One final thought. Leonardo was struck by the repeating patterns he discovered in nature in completely different contexts and at vastly different scales. The branching of down into finer and finer structures of the human circulatory system matches the branching of trees and waterways of a river delta, to describe one.
Recently, I was marveling at the graphical image of our current model in the editor. Today it looks like a depiction of a globular star cluster connected by a hub-and-spoke transportation system. That is because we are only mostly seeing just the main connections of category and epistemological strata. But with time, as each node is more richly modeled, we will see gradually see a dense web of interconnections from each node to a number (10?, 100?) of other nodes. As this happens we will gradually see another unmistakable pattern from nature emerge: the neural network. Co-incidence? I think not. Significance? I have no idea but it seems worth thinking about.

We are describing our knowledge of the world – our scientific knowledge, as a structure of highly interconnected nodes with very explicitly designed connectors which govern the flow of information between nodes deterministically. This is working really well. Why? Hard to say and we really don’t need to explain why any more the Roman’s had to explain why the Pantheon doesn’t fall down. But – just to speculate, maybe it works because the real-world is itself, at least at one epistemological level, made up of nodes and connectors – so the atomic theory suggests. A repeating pattern in nature.

In the meantime, we will continue piling one stone on top of another, building the arch that will support the bridge that our sapiens will cross over to become the world’s first AGIs.