Scientific Brain Model Chapter II

 CHAPTER II - What features carry the most significant information on the things out there, in order to solve the relevant tasks to achieve our goals?

Now, admittedly, in the simple frog's example above, our friend did not need a whole lot of incredibly sophisticated knowledge, in order to be able to get itself a tasty snack. Basically, all what it required was the correct visual recognition of a fly in its environs, plus precisely estimating the prey's exact position, as well as accurately snapping at it. Fortunately, for such a simple task, it did not matter what kind of fly it was: male or female, old or young, etc. As a matter of fact, subject to our friend's better judgement, it probably is quite as happy catching house flies, fruit flies, crane flies... Rather, it seems reasonable to think that size should be the critical factor; so that the larger the prey, the more desirable it will be and, therefore, the higher firing rates it will elicit from those neurons responsive to it. 

There are many cases, however, where tiny physical (visual, auditory, olfactory, tactile, gustative...) differences do matter a lot. Our froggy friend may not care about its prey's sex; but a farmer hoping to make business from selling eggs will certainly be concerned about the sex of his newly-hatched chicks. Fact of the matter is he will not want to spend money feeding the creature for many weeks, only to eventually realize that the thing came out to be a rooster. A city dweller will undoubtedly find male chicks and female chicks absolutely indistinguishable; but for our egg-farmer, if his primary concern is business, he will make sure to learn how to tell a male chick from a female chick, right as they hatch out of the eggs. 

Probably even more so interesting is to consider, that very often tiny physical differences are insignificant for some task, but are very relevant for some other. For example, skin color does not bear any useful information to distinguishing men from women; but is most relevant to tell the race of a person. There are even cases where two patterns are sensorially identical; yet represent completely different concepts: for instance, a live tiger looks absolutely identical to a picture of a tiger; but they obviously represent completely different concepts. If we want to understand how recognition is conducted in the brain, it is fundamental to keep these caveats in mind, since the whole point of perception is to recognize what is the source of the observed data.

 

* How do we select an alphabet of features, with which to code descriptions of the things in this world?

In fact, though not necessarily completely refute it - the previous observations certainly speak against the view, whereby a comprehensive description of the environment is first built, and from which a plan of action is then determined: If perception is decoupled from action, and no decision nor action is taken all until a full description of our environs has been completed; sooner or later we will be confronted with a situation, where the internal representation of our surroundings created by the sensory brain areas has not capture all the detail necessary to produce an accurate response. For instance, in a mountain scene of a river flowing down the slope, there will be many rocks visible along the river banks and the actual water stream. However, under normal circumstances, we will not consider each of them individually, as a body of its own; but they all will just be part of the river's global-picture concept. Now, if by any chance we need to cross the river, every rock will become relevant in its own right, as we look for any support point, where we can step along our way across. THe argument applies even more so to the sheer design principle; Namely, if an hypothetical city dweller, who has never been to a forest before, is required to learn to identify different tree species based on their leaves; he will find that Biederman's generalized cylinders shape-vocabulary (which is so very well suited to represent human-made objects) is not an appropriate tool set to accomplish such task (for only one thing, Biederman's 'geons' would not be able to support any information on the leaf's contour). Similarly, as prodigiously rich as his vocabulary was, Shakespeare would not be able to communicate effectively with an average XXI-century English speaker. In this framework, the only possible solution would be to over-dimension the universe spanned by our feature alphabet, so that it would be able to cope with any contingency we may encounter in the future; but that is obviously not an optimal approach. Thus, considering that everything in Nature is optimal - at least, as of yesterday -, we can be sure Nature found a better solution. Indeed, the optimality of an approach is always an excellent guide to find out how Nature solved some complex question. .

As a matter of fact, early experimental data showed that even the most basic physical features are learned from experience: If a new-born kitty is deprived from observing anything but vertical line patterns, it will only develop vertical-line feature detectors. Nature, however, only allows a very short period of time to learn these basic statistics of the environment. For cats, the neural plasticity (this learning process requires) only lasts a couple of weeks. If by any chance, our kitty friend travels later in its life to some other distant world, where the basic features bearing the most information (those basic features most telling and indicative of the most relevant physical differences) are different from the kind of oriented edges and color-opponency features, which are so useful down here on Earth, our little hero will find itself with no other option, than to figure things out with the basic feature-detectors set it acquired during the first weeks of its life.         

* The optimal approach is to learn a hierarchy of increasingly complex features: for instance, let us combine letters into syllables, to build words, with which we can write descriptions of the things in this world.

We can think of these basic, low-level feature detectors as the letters of some kind of alphabet developed by the brain to code the information it receives through the senses about the surroundings. However, for the same reasons, that it is not very helpful to describe a picture in terms of a massive mosaic of little oriented lines and tiny colored circles, a coding scheme based on awefully long strings of letters is likewise inefficient. Alternatively, Biederman's Geon Theory suggests that the brain extracts these simple, basic features to in turn identify complex 'generalized-cylinder' shapes in the environs; but such a description does not represent much of an improvement. Rather, it is certainly easy to understand, that just a verbal description will generally be far more useful: "There is a mug of tea on the table". Framed in these terms, it is possible to envision at least three different coding strategies:

In the first approach, a moderately sized alphabet is employed to code scenes into an awefully long string of letters. This corresponds to a depiction of the visual scene in terms of little oriented lines and tiny colored circles.

 

A second strategy would follow along the lines of Biederman's generalized-cylinders framework; namely, a very complex (Hieroglyphic or Chinese reminiscent) alphabet is used to code scenes into a moderately-long string of characters. In this case, a visual scene containing a mug sitting on a table would be described as a small cylinder receptacle, with a little, curved, thin cylinder handle attached to its side, sitting on a wide, flat cylinder board, laid in turn over four tall, thin cylinder legs.

     

The third approach would combine letters of a moderately-sized alphabet into syllables, to build words, which - placed one after another - in turn form sentences. Obviously, if one sentence is not enough to capture all the information with sufficient detail, more sentences can be added as necessary. I am sure you will not have to think long, before you can guess what is the strategy Nature chose...

Optimality is, however, not the only indication we have to guess Nature's approach. Indeed, neurophysiological recordings reveal that, as we progress from low-level sensory areas deeper into higher levels of cognition, neural responses become selective to increasingly complex sensory patterns. Since the visual modality is so dominant in the primate brain, it is easier to make out this design principle in the ventral "what" visual pathway (the neural pathway responsible for identifying "what" bodies are present in the visual scene). Thus, while neurons in primary visual brain areas V1 and V2 are selective to simple, basic features such as oriented edges and color opponencies, neurons in V4 respond to medium-complexity features, like curvature and convexity, and neural responses in the inferior-temporal cortex (IT) are selective to even more complex visual patterns, such as shapes or facial features.

  

At first glance, it may appear that the third approach lies somewhere in between the first and the second; but it is actually much more than that: certainly, it is far more versatile. The genius of Nature's sscheme resides in that it employs multiple levels of representation, building up one on top of the previous, and -crucially - getting more and more expandable and adaptable as it grows from one level into the next. A simplified way to think of it is considering just to levels of representation; namely letters and words. In the analogy suggested above, (for the visual modality) the letters are the basic oriented edges and color opponencies of V1 and V2 neural responses; whereas the words are the complex features (shapes, facial features and similarly complex patterns) signaled by neurons in the inferior-temporal cortex. Then, in the same way that a child learns how to speak, the animal brain learns to combine letters of the alphabet (i.e. basic sensorial features) to form and acquire new words with which to give names to those ideas and concepts, which prove to be more useful to establish the specifications of all the different tasks the animal carries out in its daily life.

 

* We become more sensitive - learn a richer alphabet of feature detectors - for those patterns we experience the most.

An imaginary child, who is raised entirely within a law firm, will grow extraoridnarily proficient and skilled in the kind of legal jargon generally employed by lawyers. It, for example, will know very well the precise differences between all the various types of crimes: felony, misdemeanor, manslaughter, murder, etc.. On the other hand, it will only have a very rudimentary understanding of, say, the medical jargon. Viceversa, another imaginary child, who is raised entirely within a doctor's office, will grow extraordinary proficient in medical jargon; but will, otherwise, be completely at a loss when it comes to comprehending legal jargon. For it, there will only be crimes, some more severe than others, but, in essence, all basically crimes. Furthermore, neither of the two will, evidently, have any clue of any other jargon, such as cooking, mechanical, etc.. .

 

I grew up in Spain and therefore can tell very easily between the many distinct Spanish accents existent in the different regions of Spain: Madrid, Galicia, Western Andalucia, Eastern Andalucia, Basque Country, Aragon, Catalonia, etc.. However, when I traveled through Mexico, as much as my Mexican friends would swear there is the same kind of rich variety of Mexican Spanish accents, to me they would all sound basically the same. Needless to say, my Mexican friends will not be able to hear the differences between Spain's Spanish accents either; yet, there are so dissimilar! Likewise, growing up, I barely ever came across any East-Asian person. In those days of freedom from today's hypocritical political-correctness, there was a joke, according to which, if you were standing at an airport's checking counter, watching as east-Asian tourist arrive to do the check-in, you would never know if it was different people checking-in or it was only the same guy repeating the process over and over again. It was then quite funny to learn when I traveled to Japan, that Eastern-Asian folks get the very same perception; namely, to an hypothetical Eastern-Asian child, who is raised in some small village deep in the countryside, and has never come across a non-Eastern-Asian person, all Whites would look basically the same.

* The brain develops a model of the world, describing in more detail those patterns which are more common.

  

 This phenomenon is not restricted to auditory or visual stimuli; but actually applies to everything. Indeed, it reflects the most fundamental principle of how the brain codes the information it receives through the senses. Essentially, in order for the brain to learn to optimally interact with its environment, it is paramount to build a faithful model of the world outside and, for such purpose, its statistics need to be learned. As a result, we will be more sensitive to those stimulus categories that we experience most often. As a rough rule of thumb, if, for example, 10% of the stimuli we interact with are of blue color, 30% are yellow and the remaining 60% are red; then 10% of the neurons responsible for coding color information will develop some preference for blue color, 30% will skew towards yellow, and the remaining 60% will fire more vigorously to some red color hue. Similarly, as our dear kitty friend could tell you, if, for example, 60% of the lines we perceive in our world are vertical, 30% are horizontal and the remaining 10% are diagonal at various angles; then we can expect that 60% of the edge-detector neurons in the primary visual cortex will become selective for vertical lines, 30% will prefer horizontal lines, and the remaining 10% will respond more strongly to diagonal lines at various angles. Now, you may have got really concerned thinking, that your brain only has a very short window of time right after birth to learn the most basic features of the environment; but that is only because there is really no reason to expect said basic features to change over the course of your life. Or you really have plans to travel to some bizarre galaxy sometime soon? In contrast, the neural plasticity required to acquire strong sensitivity for complex (high-level) features lasts much longer. Evidently, it is far more reasonable to think, that you may be transplanted to some distan tpart of the world and get exposed to different accents, facial features or any other complex (high-level) feature alike.

So far, I have centered the discussion on the sensory domain; but, obviously, the principle most absolutely applies to the action domain as well. Certainly, still at advanced stages of life, we are capable of learning the kind of neural motor programs responsible for the execution of new tasks and skills. Furthermore, we evidently become more dexterous and accomplished at such tasks, the more we perform them. 

As a matter of fact, the way we interact with the environment has a fundamental effect on how information is coded in our brains. For one thing, the kind of tasks we carry out on a regular basis determine the setting in which we will spend more time, and, therefore- impact the distribution of the sensory information we perceive. For instance, a forester will spend far more time in a natural setting than a city dweller, and will then obviously develop a higher sensitivity to natural patterns. However, the implications extend far deeper.