Friday, December 21, 2007

The Irrational Android

I've written the last few posts from a philosophical perspective on how the mind works, with input from some different sources and books and some of my own reasonings and reflections.

It is a very difficult thing to introspect the mind. Reverse-engineering yourself is an absurd idea. A computer might be able to register and find out about how other computers work, but it is very unlikely to gain some kind of consciousness or knowledge about the state of a single transistor in itself, unless it was specifically designed to be so (which therefore generally has a reason).

Artificial intelligence uses "neural networks" to solve some specific types of problems, which in general is a fancy way to express "pattern-seeking". Some pattern exists in some glob of data and the network is being trained to process the information and have, as its output, some new parameters or data. Sounds pretty simple. Many networks like this have 16 neurons only, yet are capable of inferring quite some knowledge. For a good example of such, read the following example on neural networks and download the example executable:

More links:

So, in the example you could see that the minesweeper only has as its input the direction to the closest mine. There is no algorithm or other function that changes the direction of the minesweeper directly. Each minesweeper has its own neural network (its brain). The output of the brain is directly used to control the side-tracks (left and right). A very simple physics engine then calculates the rotation of the minesweeper and where it goes next.

The first generation is quite useless, but then winners that picked up at least one mine are promoted to the next generation. It might create children. Eventually, after 50 generations and "natural" selection (of the fittest), the minesweepers exhibit rather intelligent behaviour with regards to the mines and there are many more minesweepers that become very efficient and picking them up, even change direction immediately to the next available mine as soon as the one in front was picked up by another.

A neural network was modeled after the biological brain, which has 10,000,000 neurons, dendrites (the input to the neurons) and axons (their outputs). Something happens inside the neuron that results in the neuron triggering an output charge or not. This only happens after a neuron receives an input charge (or various).

The most interesting questions here is whether 10,000,000 neurons are the human brain, are human thought and allow for reasoning, or are only part of this. Emotions, the way we call them and how we sometimes uncontrollably display our emotional state, suggest that it is not just a computational function. In other blog posts, I reasoned that emotions are the driving forces behind humanity. I also challenge the line of thought that rational thought really is rational in nature (void of emotion), because I recognize that most of our human actions and interactions are based somehow on emotional action or response, where it also uses some reasoning to add to the response or subtract (withheld emotions or exaggerated emotions). I think every one of our decisions, which people sometimes label as rational, are actually emotional decisions with a cover of argument. Only in the case where we reason within a factual model (science, maths, etc, which are generally highly deterministic and consistent) can we state that our reasoning is void of emotion (2+2=4 and will never be 5, whereas our human decisions in similar situations can highly differ through the concept of priorization of emotional importances, which are then argumented further as if these were intellectually considered points).

Should an android like Data from Star Trek exist, then I would not expect him to display emotional state, nor advise anyone on the best course of actions through a lot of reasoning. One could not ask Data if he wanted to go out for dinner, because he wouldn't be able to resolve the question, as he doesn't feel anything and cannot reason within the emotional domain. One could not ask him if he liked the color red, or if he wanted to take care of little androids in his life. To want and to like are unimaginable and unresolveable concepts for the android.

So... what does this mean for us human beings? The brains are a bunch of neurons and outside the microscope we see a hump of meat (humor, sentient meat):

Is it really possible that 10,000,000 neurons connected together can fool us into thinking that we are conscious beings? Is consciousness embedded within these neurons or is consciousness yet another force in our brains that uses the neurons for reasoning and recognizing patterns?

If it is true that the neurons drive us, then we are basically nothing else but pattern-recognizing machines. Everywhere we go, we quickly recognize the world around us and continuously reinforce new patterns when experiencing things never experienced before. It is basically input->processing->output for everything we do, look at, hear and so on. To ourselves, our thoughts claim that we look at absolute objects, have absolute knowledge of the world around us. That is, we recognize that a brick is a brick and it cannot be something else. But maybe the real way *inside* our neurons, how we look at the world is through pattern recognition. If we then consider "thought" as the output of the network, then it is clear that the inner working lies hidden for us. We don't have specific knowledge or awareness of the process within the network, since the output is what we measure.

This might mean that the experience of this world is highly driven through (partial?) neuron activity. You see something, it gets processed, it activates other neurons, and so on, until at the output we recognize it as something.

A mistifying factor here of course is the ability to learn. Not "learning" in the sense like artificial networks do, but learning in the sense to have an interest in understanding something. An ANN is a network that has a very specific purpose and is conditioned to execute its purpose with a training set. It recognizes only that pattern, but don't ask it to reason about something else.

Reasoning can probably also be expressed as the activity of comparing pattern recognition programs with one another in the hope to reuse or pre-initiate a new capability of a new network.

Another limiting thing for androids might be that I do not expect them to start asking questions. Would Data ask anybody else about a particular event? He can be given a goal to execute, the result can be right/wrong or informative, but will Data ask questions from the environment to enrich his own knowledge? Or will he actively seek this information on the Internet?

Looking at a developing baby to a toddler, there are many interactions with the environment. Given a clean slate, the baby starts to develop an interest in the environment. When it is born, it cannot focus the eyes or move the muscles through coordination. These can probably also be seen as "patterns", where the sight and hearing are paired together and then compiled in a network, where the output is directly used to control the muscles. Research here in the area of clock frequency or how fast our movements are adjusted based on changes would be interesting.

In further development, children are shown children's books and we point at pictures and say "zebra" or "tiger". At some point children get it. Then ask a child..."Can lions walk?" and the child uses reasoning to find this out. Uncertainty is a feeling one can have about the validity of a certain answer. This has a direct relationship with the strength of one belief. If a belief is formed by the strength of the trigger of a set of neurons, then you could say that it is also how strongly a pattern has been recognized. If a lion has legs that bend backwards and looks somewhat like a cat or dog, and things that have legs can walk and cats and dogs do walk, then with a good amount of certainty, it can be said that lions walk too. Now consider the parent... there are many different little bits of knowledge that we pair together in our reasoning that the parent uses to teach the child (with that usual mother-kind-of-voice): "The lion has legs and manes and big teeth. It can bite and is dangerous".

This pattern recognition, if it is at the source of intelligence, then also explains why we are so subject to categorization. By forming larger categories, we collect truths in the same basket and then test other things against that rule. That greatly reduces the need for storage space in the brain and makes it possible for us to make assumptions about things never encountered in this world.

Problem solving skills are the next step then. Given a particular goal, we can think and think and come up with a solution to meet that goal. We could say that we have the ability to 'imagine' things we have seen happen in real life and then try to replicate that same event, behaviour or property with other means. Problem-solving highly depends on the ability to create different hypothesis and to recognize (cross-pollinate) ideas from different analogous areas.

It sounds as thus, if the brain has very powerful pattern recognition abilities in the brain, but it is helped and supplemented by additional capabilities that use this network as a tool. Imagery patterns, auditory patterns, behavior patterns and so forth. When studying the child, one can clearly see that it is a lengthy process to get our neurons in order. Learning takes a long time to finish and even then, us humans do not develop in the same way and do not exhibit the same behaviour. Each one of us is unique. Our ability to deal with noise that clouds the pattern is amazing.

An android might actually have some advantages in learning. It should be possible for a computer to demonstrate internal states of neurons and visualize them, or execute "what-if" scenarios (as it is computed and won't have a guaranteed effect on the state or quality of the network), so that we can steer the development of the network or understand it better. One can also imagine a desktop that demonstrates the processed elements with interactive handles to improve the network, the same way a parent would point at things in pictures to explain why certain associations are true and how these same properties in other pictures, which may look somewhat different, also demonstrate the same behavior or constitute the same thing.


Don said...

Some thoughts.

"I also challenge the line of thought that rational thought really is rational in nature (void of emotion), because I recognize that most of our human actions and interactions are based somehow on emotional action or response, where it also uses some reasoning to add to the response or subtract (withheld emotions or exaggerated emotions). I think every one of our decisions, which people sometimes label as rational, are actually emotional decisions with a cover of argument.”

Perhaps you would allow that most of our actions arise from the simple physics of moving our bodies through a highly asymmetrical physical environment. Further, you would probably accept that the fundamental need for human movement is driven by the general imperative of all living systems: balancing the equation of Calories in with those of Calories out.
As I understand it, the majority of our physical movement is governed by neural networks that do their job without “calling home”. As far as deciding what to put in our mouths, the simple discrimination of food items from non food items is likely done by simpler neural networks but, just as clearly, there is a higher level emotional roulette which may finally determine whether it’s fruit salad or chocolate cake on the fork.

The ai-junkie link and his introduction to neural networks is the most accessible and concise I have come across. (by the way, he uses the figure of 100 billion neurons in the human brain rather than 10,000,000)
Looking once again at the familiar schematic of a neural network I realized I had not thought much about the physics of networks, that the work of pattern recognition is highly negentropic, that is, the multiplicity of information “vectors” in the data set may be reduced to a single binary state space. Complex phenomenon are abstracted into simpler “tokens” and surely this must have relevance upon the argument as to whether the stuff human thought is symbolic in nature. Clearly the real work expended in pattern recognition must find its return in the larger equation of sustainability of the organism.

Gerard Toonstra said...

One of the things to remember is that the processes in older elements of the brain, those elements also found in animals and which react instinctively, are also active. We're not always reasoning or recognizing patterns consciously. I'd be interested in learning what "calling home" means.

With regards to ANN, it is a simple model and I have sometimes difficulty to imagine that the binary state can 'represent' any such complex situation as we are conscious of in a given day.

Consider the signals that we are processing and that the human brain is biological in nature (not digital like computers), and it is subject to both electrical signals and chemical signals... one can only imagine what kind of signal flows between the neurons.

Besides this, the ANN is in general a feed-forward network. That is, it is excited at the inputs and always moves forward. There are other networks that apply feedback, have recursive loops and where the output may be applied as an input into the next cycle.

The last thing that may be different is the clock cycle. A computer is triggered by a crystal that a high number of seconds emits a trigger signal. This crystal is at the heart of the CPU and drives another cycle of processing. If we consider the state of a biological brain as given at some point, but we claim to be in full consciousness at any point in time, how does this apply to the analogy with computer brains? Do biological brains have clock cycles at all and if not, is this the secret behind full consciousness and may this be some kind of key to a more fluent kind of consciousness?

The reason for asking this is also with regards to when the output should be measured. Should this be after the full network has been processed (re-calculated), or should it be possible to measure the outputs at any point in time, or should we only measure it after any of the outputs have changed?

Gerard Toonstra said...

Your final paragraph contained a very important essence (phenomenon deconstructed into simpler tokens). This might be at the heart of all our processing, although we might not be aware that it happens. Memory however does give us cues. Details in our memories fade away, but specific emotional events or vague elements do not. Even if the details cannot be immediately recalled, we do recognize the details when we are presented with those (as if these were 'grayed' out somehow to fade into black, but come back in living color).

Considering how I for example remember a certain scene after a quick glance, I can only redraw rough lines of that scene, rough shapes in black and white without exact color. Maybe that's how we are able to very quickly move around in fast situations, by reducing a particular scene to very basic elements so that recognition is quicker (add the element of 'expectation', what you'd expect to find somewhere) and you have a machine that can very quickly orient itself.

This is perhaps one explanation why we marvel at (fine) art and take long to look at it. People that appreciate art do not look at the big, rough picture, they totally go up in all the details. Paintings are often characterized by specific brush strokes, colors or shade colorings. Audiophiles claim to hear specific differences in timbre of high-hats inbetween loudspeakers and amplifiers.

Don said...

By “without calling home” I meant just what you suggested: many actions are taken without involving the conscious mind. One figure I have seen given is that the un-conscious mind handles about 40 million bits/sec while the conscious mind processes at about 40 bits/sec. Clearly some of those 40 bits/sec must be binary “tokens” conveying more than their bit count would suggest. Just as a coin outside its country has no currency, a bit in isolation has no meaning whereas; within the context of a coherent network it confounds any information metric. "Music is what happens between the notes, not the notes themselves." (Concert pianist Karina Sabac)

The key strategy of the mind may be pattern recognition, its ability to say, “This is like that”. Often the “that” is not present in hand but somehow stored in memory. As you mention the brain has different clock rates and somewhere I read that memory is refreshed by regular waves at about 40 cps. Understanding the brain may be like trying to map the outline of all the continents by analyzing and extrapolating the wave patterns in a small patch of ocean.

From the little I know I am awed by the complexity of the brain. Somewhere I came across an estimate (don’t have a reference) of the number of possible synaptic connections in the brain and it was of an order of magnitude similar to that of the estimated elementary particles in the universe (10^80). That is hard to believe and I wish I could confirm it.