What is Intelligence: A pass at creativity

A PASS AT CREATIVITY


The human mind viewed in a certain way consists of a simple table lookup system that maps stored patterns to a menu of possible actions, and while many may frown at this simplistic view we must realize that it is by looking at the simplest possible model of things that we can really understand them. When we scrape away most of the details we can see the barebones implementation, understand them and then start adding sophistication.


The basic patterns that make up the substratum of the human mind are actually obtained from the natural environment that a particular human is immersed in but the human mind using the synthetic capacity for imagination actually creates new patterns from the raw base of patterns supplied by direct experience.

This synthetic capacity is what is known as creativity and is probably done by some kind of Generative Adversarial Network (GAN), which generates fake data that mimics some aspect of what has already been perceived from nature and remixes these data to create something that doesn’t exist but which contains patterns that are preferred by the entity.

Juergen Schmidhuber has some very sophisticated definition of creativity as seeking novelty, but I think that his formal theory of creativity is more of a product of his own creativity and extreme intelligence than it is related to anything the brain might actually be doing.

The basic algorithms of the brain are not as sophisticated as might think, the sophistication we see in nature is actually an artefact of our own intelligence which is actually based on extremely simple stuff. Simple algorithms like the very simple functions that combine to enable us to recognize images make up the majority of the kinds of stuff nature does at its lowest and most primitive levels.

It is only when lots of these simple things are connected and combined and data flows through them and is transformed in numerous ways that we have the complexity and sophistication that we see in nature and in the human brain.

When I say that imagination which is the foundation of creativity is carried out by GAN like systems, I do not imply that there is a GAN in our human brain. What I am pointing at is that the system for doing that might be working like a GAN or even the kinds of algorithms that do Google Dream or Magenta or even Stephen Wolfram’s cellular automata, Mandelbrot’s fractals or some other niche algorithm that I have not yet had the chance of exploring or some other stuff we will come up with in the future. All these various implementations of algorithms that could mimic what we know as imagination and creativity are all just specialization of what could really be going on, and what is going on might be extremely much simpler than any of our implementation and much of the sophistication we see in creativity might just be an artefact of this very basic algorithm.

Seeing the creative process from the highest point we could just see that creativity generates information, and we know that the human brain has an enormous database of information it has observed from its environment so therefore it feeds on that data to create new data that differs from anything it has observed before in the highest scenario of creativity.

It is common for very smart people to point to the works of great composers like Mozart or Picasso when they want to describe creativity. But creativity could be as simple as scratches made on a cave wall with no intent to impress whatsoever. Focusing on this high points of human creativity might deceive our minds that creativity is something super sophisticated but if we come to see it from the foundations, those high creative works like those of Leonardo Davinci are no different than the activity of some class four cellular automata evolution. The only difference is that the Monalisa painting was done by a human with a lot of cumulative history embedded in it while the cellular automata is just some process executing an algorithm which is producing output above some threshold of statistical randomness but at the foundation, both of them are somehow creating fake data, that looks like some true distribution no different from what a GAN will do.

The point of what I am trying to say is that creativity is not some very high novelty seeking aspect of the human mind, although at its highest expression it is what Schmidhuber describes but trying to build a creative machine, which may eventually become an artificial scientist by trying to build a novelty seeker is just doing what the GOFAI guys were trying to do with knowledge in the Cyc project and others. If we are to learn anything from them it is that intelligence, its expression as rational action, knowledge generation or creativity is not based on some sophisticated process but some simple almost stupid basic principles that do not look like intelligence at all.

It is when these simple things are interacting in simple ways in large numbers that we have what we see as a complicated system expressing high-level stuff like creativity. The biggest discovery of our times is that we can do image recognition with simple things like convolutions and matrix multiplication on input data, this is the ultimate revelation of this 21st century in AI research. My advice is that we should not get stuck on these low hanging fruits, there could be far more effective ways of doing what we are currently doing with convolutional neural networks and we should seek those simple processes by searching in nature, our own creativity or elsewhere. We should not get stuck on our success, it is obvious that convolutional neural networks work, but the data and power requirements are too much. In the future training and inference will not occur on power hugging data centres, it will occur in our pockets, requiring little data and power.

We have entered the seashore of discovery with Convolutional neural networks, and other deep implementations. Let's not get stuck, we have only held a grain of sand in our hand there is so much to explore by going back to basic research, rocks and even sea are abound.

When required to act in the world the mind-brain system generates a large swath of possible actions by combining already existing patterns that have been generated using the generative algorithms of the human mind that work like GANs, RNNs, etc. and passes the generated structures through a mental DeepRL (Deep Reinforcement Learning) like system internally. After running simulations subconsciously the system generates an action in the world, using its internal rewards system the mind determines if the results of its action are beneficial to it or not. In typical DeepRL all the effort of the agent is directed toward either making a move on some kind of board game or performing some grasping or moving action by a robot. But the mind’s internal system of reward an punishment might be much more subtle than that.

When creating a piece of art it is still this same reward system that is responsible for modifying the path of development of the artwork, eliminating developmental paths that might not generate much pleasure (reward) and emphasizing those that generate pleasure. Most of this is occurring internally in the subconscious so the human does not make the action in the physical world all the time before modifying it.

If an artist is doing a painting or some music or some other work of art, its internal reward system constantly rejects proposals from the generative system after running it in some internal simulation and seeing that it did not like it, that is, it did not generate as much pleasure as it anticipated. Most of these actions are eliminated and only those that have the capacity to induce much pleasure in the agent are left. Out of all those, there might be some candidates that are so close in how much pleasure they generate that the agent cannot determine from imagination alone which will turn out to be the most pleasurable. So the agent resorts to acting them out in the world.

This might be seen when we draw and erase because what we drew was not satisfactory or did not evoke much pleasure as expected. In the early times when we are developing creative skill, it might be that our limbs are not yet tuned to very high levels of precision so we make a lot of mistakes due to undesirable movements. In this case, our internal system might have a single powerful proposed structure that it wants to execute through its limbs but the limbs owing to poor development are not able to execute it perfectly, so it tries repeatedly until it attains a good amount of precision. Assuming the limbs have been trained through practice, the goal will then be to bring out the expression that generates the most internal pleasure not necessarily avoiding poor physical execution by the limbs.

In the case of the untrained limbs the mistakes are due to poor physical execution, but in the case of the well trained the mistakes are due to not being able to choose the amongst a bunch of competing candidate creative proposals which one will lead to the highest amount of pleasure or reward. In this scenario, the only solution is acting them out in the world and judging which turned out to look better.

In any case even after acting out all the proposals in the world, and especially when there is high fidelity in the limbs of action, it is still not clear which is preferable. It is at this point the agent chooses randomly over the set of manifested proposals.

Every single output of our minds no matter how abstract, even as abstract as pure mathematics has its basis in input derived from our natural environment. While the primitive man might draw actual scenes that happened on cave walls, the abstract artist might convey emotions with images that might not mean anything to anybody but themselves and the pure mathematicians might deal with relations that do not obey any physical law. At the bottom of it all, at the lowest level, all these are generated by very simple algorithms running on brain cells and are all still equivalent. And if the primitive man, the abstract artist and the pure mathematician are all running the same algorithm on the brain cells and connections, all their expressions are rooted in information obtained from nature.

Any action we take whether it is moving from one point to another, solving a sudoku puzzle, drawing some image, producing a sound, or generating a mathematical proof all generate information which can still be feedback into the pattern recognition system again and end up as new patterns of integrity that are stored in the brain. This is actually what is going on in our minds and because the brain is capable of creating new patterns and actions beyond what has been input to it directly from the external environment, it is hard to trace the origin of every piece of knowledge (pattern integrity) or action sequence that the human mind generates and thus building synthetic intelligence which many people are currently engaged in because of the benefits that could accrue if success is achieved has been horrendously difficult.

It's like we are in a room full of mirrors and previous GOFAI researchers were trying to capture what was actually projected on the mirrors because they didn’t know that most of what we see as the effects of our mental processes are just like the images projected on the mirrors. If the arrangements of these mirrors are altered like in a kaleidoscope what we see will be modified in ways we may not be able to understand. For the first time with neural networks, we have been able to grasp on one object at the centre of the room and we are no longer chasing the illusions generated by the mirrors. What we have to realize is that the object we have grasped, neural networks, is just a specialization of more generalized network structures and rather remain stuck at the trunk of the tree of intelligence we should go deeper towards the roots.

At least we are now dealing with the trunk of the tree of intelligence it is only a matter of time before we get to the roots, then the seeds and from there we can invent any kind of tree or plant we want, human, animal or alien intelligence.

Comments

Popular posts from this blog

Software of the future

This powerful CEO works remotely, visits the office only a few times a year

Nigeria and the Computational Future