The meaning of consciousness: We are conscious of something, if we can express it with words.

The meaning of consciousness Part 1: Extracting the complex features which identify our conscious concepts. 

Concepts and conscious (natural-language coded) knowledge: (conscious) concepts are the stereotypes we form to use as sequences-of-actions- building blocks, so that it becomes feasible to elaborate complex plans for the achievement of our goals.

* Unsupervised learning is followd to learn low-level feature detectors of those patterns, which we find more frequently and most faithfully describe the environment; whereas supervised learning is followed to learn those high-level features, which are better suited to code the tasks we execute to achieve our goals.

At this point we enter in the field of supervised learning. Until now, I have been discussing forms of unsupervised learning in the brain. However, generally, in order to be able to achieve some goal, some form of supervised learning will ultimately be necessary. The role of unsupervised learning is to highlight and squeeze out the most relevant information about our surroundings; however, in order to achieve some goal, eventually it will be necessary to settle on a course of action. Supervised learning is then the process whereby the system learns what is the most appropriate course of action in a specific situation. As rational as we enjoy thinking we are, fact of the matter is the immense majority of our learning occurs through trial and error (after all, it is generally the most effective method): Given a specific situation, some action is taken. If a positive result is achieved, we will be more inclined to repeat it the next time we are faced with the same scenario. However, if the outcome is negative, next time we will rather try some other course of action. The crucial question then becomes to accurately determine, when does the exact same scenario present itself again. 

* Supervised learning: let us assign a name and create a specification of complex features for those scenarios most significant for the attainment of our goals.

In order to be able to identify specific significant scenarios, complex feature detectors are acquired in higher-level brain areas. These complex features will signal the exact identity of the current scenario. Indeed, it is very useful to think of this learning process as the assignation of names to specific prominent scenarios. Whenever we reach some major success, we backtrack to the original state, that constituted the point of departure of the successful course of action. This starting situation is given some name, so that Next time it presents itself again, from its name we will know how to act. 

For instance, a young man looking for an adventure with some girl, will quickly learn, that he better focus his efforts on a girl, who exhibits a favorable disposition. Clearly, if she throws some dirty look at him, he will only be wasting his time. Needless to say, the best indication that she has some interest, is if she smiles at him. Consequently, the shape her mouth takes becomes the (complex) feature, which will signal the scenario where it may be a good idea to move forward and "attack".

Consequently, as it turns out, the kind of tasks we regularly carry out greatly determine, what feature detectors develop. As a matter of fact, this is particularly so for the more complex (high-level) features. Indeed, while basic (low-level) feature detectors emerge through unsupervised learning to reflect the statistics of the world outside; the higher up a feature detector resides in the representational hierarchy, the bigger influence we should expect the kind of supervise learning required for the acquisition of a skill must have hadin its development. Now, if the tasks we perform most often are a decisive factor to determine what features in the environment to look for and pay attention to; then it becomes even more evident that it is a poor design to decouple perception from action. In point of fact, it only makes sense to shape our analysis of our environment, based on how the acquired information is going to be used. For instance, if one is going to work as a lawyer, it only makes sense to grow one's vocabulary to gain a perfect command of the legal jargon. Likewise, if one is going to work as a medical doctor, it only makes sense to grow one's vocabulary to gain a perfect command of the medical jargon.

* Since the brain's aim is to understand how things work, the meaning of a 'thing' is determined, not by how it looks or feels; but by how it relates to our goals and other 'things'.

Here also we can see the main failing of current Artificial Intelligence systems: they are only useful for those problems for which they have been trained. If we consider again the example above of the imaginary child, who is raised entirely within a law firm, it will be very proficient in the practice of law; but it will otherwise be totally hopeless at anything else. The same could, obviously, be said of a child, who is exclusively trained in the practice of medicine: it will grow to know everything about illnesses and their corresponding medical treatments, but will be otherwise ignorant in any other domain. THe problem is not just that our legal and medical geniuses will not understand any jargon other than their own; but they will not even be able to fix themselves a sandwich, drive a car, let alone fix a flat tire. Consequently, since nothing in this world takes place in a vacuum, our friends' tunnel vision will limit them even in their field of expertise. Clearly, how is a lawyer going to be able to make a good case in a car-accident lawsuit, if he not only ignores everything about cars, but does not even know how to drive?  

Undoubtedly, it is far wiser to first get a general education, and only once one has a global understanding of the world, specialize in a specific field. That is, however, not how Artificial Intelligence systems are developed. Rather, much like our imaginary lawyer and medical-doctor friends, current AI systems are, right from their very inception, designed to carry out one specific task. 

With today's technology, it would not be a big deal to train a system to recognize whether a face is smiling or not. However, if we next want the system to identify the face's gender or race, it would not be possible to reutilize much of the logic implemented for the recognition of smiles. Basically, in order to resolve each of these tasks, the program will look for completely different features, bearing little relationship with each other. In sharp contrast, a child would learn concurrently to perform all three of these tasks and many more. As a consequence, as previously described, a hierarchy of features will form. In the lower levels of the hierarchy, unsupervised learning will yield a comprehensive set of feature detectors, specially conceived to squeeze out all the information necessary to form a detailed description of a face. These low-level feature detectors then serve as information-rich building blocks upon which more complex features in higher levels of the hierarchy can build up. These other higher-level (intermediate) features will be optimized to form the basis, from which a final group of feature detectors at the top of the hierarchy can easily determine a solution to all the various problems the system is faced with.    

 

Language comes here again really handy to illustrate, why simultaneous learning, in all the different tasks performed by the brain, leads to a seamless, optimal set of feature detectors perfectly suited to solve them. Let us consider how we should refer to the machine people use to make photographic copies of any sheet of paper. One straightforward option is to call it 'photocopy machine'; yet some folks find it more convenient to refer to it as "a xerox". The obvious advantage of using 'photocopy machine' is that it does not require any additional learning; but, from that name, anybody can easily figure it out, even if you have never seen any such machine before. Xerox, on the other hand, is not as self-explanatory. Moreover, if - not unlike a newborn baby - the AI system has not acquired some basic vocabulary, or simply does not have any general understanding of the world, and, consequently, does not know what paper is about, what is a machine, what is a photo or what it means to copy something; then it is certainly going to be quite a process to figure out what is this "xerox" thing.

 The key consideration here is that concepts get defined by how they relate to other concepts. For instance, a table and a desk are very similar things, in that their purpose is in both cases to serve as support to other items. However, whereas one would normally place tableware on a table, office material are more common on a desk. Once more, in the language domain it is particularly useful to consider how concepts relate to each other. Each noun can serve as agent, direct object or indirect object for certain verbs, but not for some others. Animals can eat, reproduce, sleep, love, kill, etc. ; but they would normally not crystalize, evaporate, be written, be sown, etc.. Similarly, adjectives make sense with certain nouns, but not with others. Finally, verbs likewise have their own rules on which adverbs are suited to qualify them, as well as which nouns they accept as agents, direct objects or indirect objects. Interestingly, it would be possible to conceive the meaning of any word based on these rules.

       As a matter of fact, if the whole purpose is to optimize our interaction with our environment, it only makes sense to build our understanding of the world based on how each concept functions in relation to everything else. Let us reflect on what it takes to solve the following questions: Is Tony crying? Who is Sergio talking to? Is Angelica smiling at me? Is David sleeping? Where is Maria looking? If we tackle these tasks in isolation, we will get a separate set of feature detectors for each of them. However, if we address them concurrently, seamlessly going from one to another; we will find that the eyes receive special scrutiny, as a significant group of feature detectors develops around them. Indeed, a detailed analysis of the eyes should extract decisive information towards the resolution of all five of the above tasks. As a matter of fact, it would almost seem like the eyes have life of their own. Undisputably, the eyes are more significant than, say, a patch of hair or skin.

* Those things, which we find are significantly related to the achievement of our goals, get a name and become concepts.

As it turns out, as we interact with our environs and act upon the bodies around us, we slowly discover what bears information and what does not. In other words, what is meaningful and what is meaningless, what gets a name and what remains nameless, what is coded and what is left out. For instance, whereas we employ specific names to designate vertical lines and horizontal lines, other degrees of orientation are simply lumped together under the term 'diagonal lines'. Clearly, vertical and horizontal lines are of special significance: while horizontal surfaces offer a safe ground on which other items can be placed without risk of sliding, tall structures will not fall, so long they stay vertical.

Indeed, there is a reason why some people find it more convenient to refer to photocopy machines as 'xerox': While 'photocopy machine' can mean several different things, 'xerox' is very specific and precise. Thus, if 'xerox' machines are significant enough, that everybody understands the term; then it may just be better to say 'xerox' and everybody will know what exactly you are talking about.

Yes, at the end of the day, all what really matters is that everybody understands what the thing is about. This, however, is exactly what today's AI system fail to accomplish. For instance, an AI system may learn to differentiate between women's faces and men's faces, and later between women's voices and men's voices; but it will never develop any concept of what is a woman and what is a man. In effect, all what the system will know how to do is to classify images between, say, category FX and category FY, and classify audios between , say, category VX and category VY. Yet, come Valentine's Day, do not expect the system to give you any kind of advice on, what kind of conduct to follow or how to go about it.

* If there is anything such as intelligence, it is the brain's ability to develop an accurate model of how things work, which will in turn guide the pursuit of our goals.

If a useful definition of 'intelligence' is the ability to learn to optimize one's interaction with the environment, in order to achieve one's own goals; then it would be reasonable to conclude, that our 'faces and voices classification' AI system is really dumb, since it will never be able to figure out anything other than classifying faces and voices. Given their undisputable prowess staying alive and thriving in this wild world, for all intends and purposes (other than the classification of human faces and voices), even a cockroach would be far smarter and more accomplished.         

Again, the whole purpose of perception is to support the recognition process, and the whole purpose of recognition is in turn to fuel the optimal interaction with our environment. In other words, the whole point is to find out what distinct bodies are there in our environment; where, from all words, 'what' is the key one here, as it refers to the meaning, that is all the information describing how those bodies function and how we can interact with them, in order to achieve our goals. We can therefore see that the end of recognition is not to assign some label such as FX or FY, but to deliver a key; namely, the key which unlocks all the information about the corresponding concept. Indeed, when our froggy friend looks at a fly, it does not merely see a six-legged cylinder, with two wing tapered ellipses on its back and other two tiny antennae poles on its head. Rather, it gets the idea that the thing, if captured, is going to satisfy all of its belly's desires. Similarly, when we look at a decorated christmas tree, we do not merely see a treetop, with lots of flashing lights, shiny balls and little figures of shooting stars, bells, bearded old men, reindeers, sledges, etc.. Rather, we perceive a decorated christmas tree, with all the meaning such concept represents: when do we put Christmas trees out? what mood and spirit is supposed to be predominant during that season? Where can we expect to find our Christmas presents?

We can therefore conclude that the learning process consists in building such a model of the world, where the brain maintains all the information it has acquired on how things work; so that the organism can then pursue its goals more intelligently.