The Free Press Macmillan, Inc.
A Book Review by: Paul Harris, OD
For more than half a century a major segment of the profession of optometry has been guided by models and concepts of vision that emerged from the writings and presentations of Dr. A.M. Skeffington, to whom this meeting is dedicated. From the beginning there was recognized a fundamental difference between earlier conventional wisdom or optical models of vision and that of Dr. Skeffington. With little hard data or exact knowledge of how the brain works Dr. Skeffington put together a conceptualization of the sub-processes that must be present for vision to emerge. Most of us are familiar with the four circle search model that has been used for years as a way to communicate to others the richness of this concept. For much of the time, since it was conceptualized, those who have and who continue to follow the model to help their patients have done so with little absolute pure research backing up the model. A belief in the model and concept was established and solidified as each follower discovered that the model empowered the optometrist to more profoundly and more positively affect the lives of the patients that they served. Dr. Steve Cool, currently at Pacific College of Optometry, a neuroscientist, has asserted for years m his presentations the fact that as more and more pure science research is done in the fields of neurology, psychology, biology and many other related and sometimes seemingly unrelated fields, is done, that support for the behavioral concept of vision has been and will continue to be
In other words, in the beginning the black box which includes the “how’s” and “whys” of vision was relatively large. In the beginning we new what the input conditions were and we observed output. Recall the statements: Vision is motor. Vision is output. Vision is and emergent. AL a high conceptual level the links had been made very early by Skeffington and his associates. Whether he knew at the physiological level how the brain and neurology supported his concepts and ideas I do not know and at this point in time cannot be verified. The critical point is that the link was made between input and output at the level which a clinician could understand and which was modifiable through the use of lenses, prisms and visual training.
A new cognitive neuroscience called Wet Mind, which is also the name of a book by the authors Kosslyn and Koenig, is bringing a more rich understanding of how the neurology and physiology supports the visual process as we know and understand it. The authors of this book clearly walk us through not only their current understanding of the field but also how they got to this level of understanding. They clearly demarcate that which is speculation and that which has already been demonstrated. This book has helped me understand this new emerging science. The more I read and the more I understand, the more I know about how the concepts and notions that have driven behavioral vision care are supported by the hard data being collected in research labs all around the world. In the rest of the paper I will attempt to share the highlights of the information shared in this valuable book, “Wet Mind.” The term “wet mind” is a relatively new term. For years researchers in the field of artificial intelligence (Al) have been working on neural networks and attempting to get these electronic networks to mimic certain aspects of human thought and human behavior. Workers in this area used the computer as a metaphor. When running a program on a standard digital computer, even one simulating a neural network, the program runs relatively independently of the specific hardware it is running on. The analogy would be to analyzing thought or problem solving abilities without regard for how the brain works. This branch of investigation which does not take into account anything about the brain is called “dry mind.” Central to the Wet Mind concept is that “the mind is what the brain does.”1 In order to characterize wet mind it is said that: “a description of the mental events is a description of brain function, and facts about the brain are needed to characterize these events. “2 Thus, the search for understanding must match with and work with the brain or as others have called it the “wet-ware.” In computer lingo one talks about hardware and software a separate entities. The hardware is the physical machine and the software are the stored electronic instructions that collectively are called a program. In the brain these two merge into one entity called wetware. The nature of a neural network is that the program or the software is stored in the manner m which the hardware is wired. Therefore, they become one and the same.
To help the reader understand neural networks Kosslyn and Koenig devote a chapter to “Computation in the Brain.” A very rich model is developed using a hypothetical neural network which consists of three layers of octopi in a tidal pool.
The three layers of octopi are arranged so that when the tentacles of the octopi in the lowest level are brushed by something that octopi squeezes all the octopi at the next higher level to which he is attached. In turn the 9ctopi in the second row who got squeezed themselves squeezed their tentacles that were connected to the highest layer. What can be noted over time is that as different numbers of swimming fish would brush up against the tentacles of the lowest level of octopi that different patterns of third row octopi waving their tentacles in the air would arise. In a very crude way the octopus network was reporting to the circling seagulls in the area the density of fish below in the tidal pool.
This simple system demonstrates the concepts in a feedforward neural network. All the elements connect in one direction only. The flow of information is in one direction. The following figure from “Wet Mind” shows a schematic for such a feedforward network. Note the connections and the weights associated with each connection. The number on the pathway corresponds to the relative amount of squeezing that the octopi would do when activated himself.
The key to understanding how the network learns, stores information and processes subsequent stimulation is in understanding how the relative weights of influence shift over time. In general, the more frequently a pathway is activated the stronger the weight becomes.
The less frequently the pathway is activated the weaker the weight and in fact some can become inhibitory. A negative or inhibitory weight can cause the octopi to which it is connected to need more than the normal excitation from other sources before he will wake up and decide to start squeezing those above him.
How does a network learn or how does the weighting change. For this to happen there must be some feedback to the network. In the case of the octopi this might have occurred in the following way. When the seagulls get an accurate picture of the number of fish below they swoop down and catch a snack. They might reward the octopi by dropping some of the uneaten fish down to the octopi. If the octopi report that fish are present when indeed none are present the seagull swoops down, catches nothing, and in the end does not reward the octopi. The seagull learns to ignore that response. Overtime the octopi network will alter the squeeze patterns to feed itself.
The authors later expanded the octopi network to show that it could perform two related tasks. When the tide was low and not fish were around a group of swimmers would come and play with the octopi. Different characteristics of the different swimmers in terms of how they interacted with the octopi would fire off different tentacle waving patterns in the upper octopi. Since some swimmers brought snacks for the seagull and others didn’t, over time the seagulls, through feedback to the octopi in the form of some dropped snacks, altered the weights so that the seagulls could actually tell which swimmer was in the water based on the tentacle waving pattern. In essence, the octopus network was identifying people! In the end all the network has done is to pair an input condition with an output. This is a quantitative association which is a kind of associative memory. The interesting thing demonstrated is that once the weights are set up that the network can be used to process different types of data which might require similar types of input output relationships. I quote from Wet Mind a section which helps to understand this a bit better: “The pattern of weights established on the internal connections of a network often serves as a representation.A representation is something that stands for something else The pattern of excitatory and inhibitory weights on the connections represents the combinations of features that identify the individual people. These weights store the information about the people that allows the network to identify them, and in that sense represent the people in their absence. If we think about a network that recognizes objects more generally, the patterns of weights in the network will be representations of different objects; without the right weights, the network cannot map the input to the proper output.”3
I have illustrated a simple type of neural network. It is important to understand that there are many types of networks and that the networks in the brain are far more complex with more layers and far more interconnections than has ever been drawn or modeled. In many network types the. feedback loop is built right in with specific connections from the output layer going directly back to the input layer.
The computation performed by the network is a simple association game. The authors talk about the difference functions and a Function. When spelled with a capital “F” it represents an overall process or ability. When spelled with a lower case “f” it represents sub process which when combined form the high level Function. The work done by functions is to map sets of input to different sets of output. They do this by following the rules which are how they are wired including the relative weighting of the connections. The output is utilized or interpreted by the larger system which is the Function.
The importance of this is that the field of Wet Mind researchers see the brain as made of many small neural networks which are performing functions. Groupings of these small function networks are joined in such a way to form larger networks. In essence the output from a number of small function networks becomes the input for the large Function networks. It is interesting to note that many of the small function networks feed their output to multiple large Function networks.
In some instances it is appears that a small function network may have originally developed as necessary to perform some large Function which is or was necessary for survival. In a number of cases later developing large Functions, as example the authors discuss the process of reading, have organized themselves and opportunistically used some small function networks which were not specifically organized or trained for the new emerging large Function. They just happened to already be doing that which was needed and got recruited for the job.
The field of Wet Mind researchers generally start by doing computational analysis. This is a logical analysis of the information processing that is needed to produce a specific behavior. From this information they attempt to build neural networks, train the networks and then see if the networks behave the researchers expect. The process repeats over and over until the network, or in most cases, complexes of networks strung together fairly accurately is demonstrating the original behavior.
The Five Principles
The early work in Wet Mind has led to the discovery of five principles. These are: Division of Labor, Weak Modularity, Constraint Satisfaction, Concurrent Processing and Opportunism.
The following are short explanations of these principles.
1. Division of Labor: Neural networks can be used to perform similar types of mappings. However, certain types of mappings are incompatible and cannot be meaningfully performed by the same network. Certain situation require that a job be divided between two or more networks or groups of networks in order to solve the problem. In the are of vision this leads to the understanding that there are actually separate Functions for spatial location information, where is it networks, and for object properties or what is it networks.
2. Weak Modularity: “Individual neural networks are not independent, discrete ‘modules’ within a larger system. The principle of weak modularity has two facets, which pertain to functional relations among processing subsystems and the localization of networks in the brain.”4 The idea here is that a particular site in the brain may appear at some levels, particularly at the small function level to be well localized in the brain and related to other small functions that subserve the same large Function. However, the large Functions, which are the behaviors which the organism exhibits is not sited in a specific location in the brain but may be exist only as a total state of the brain as a whole. Large Functions are not specifically localizable.
3. Constraint Satisfaction: The concept here is that the brain is capable of dealing with lots of different things at the same time. Each of the separate things may have a small constraint associated with it. The analogy was made to how furniture is placed in a room. An example of a constraint is that a bed with a rickety headboard is usually placed against a wall and a small side table might be placed next to the bed and a couch with a missing legs supported by books might need to be placed against a wall to be stable. The specific constraints of each piece of furniture are small. However, in a very small apartment with limited wall space the weak constraints of each piece of furniture may combine to allow the furniture to be arranged in one and only one way to satisfy all constraints. It turns out that the brain is excellent at solving these types of problems.
Another example of this is what they term “coarse coding” and we use it to code colors. They point out that we could have developed separate cones for each wavelength that specifically encoded the color of an object. Instead we have three types of cones. “Coarse coding is a way of exploiting the fundamental idea of constraint satisfaction; each input is only a weak constraint, and is effective only when multiple weak constraints must be satisfied at the same time. “5 The three cone types are red, green and blue. Although they have these labels they each actually respond over the entire spectrum. However, they have different response curves which overlap considerably. Coarse coding is the method of using the degree of overlap in responses from units (nerves) or networks that have different sensitivities to specify precise values.
4. Concurrent Processing: All networks are always turned on and working. Networks may process in parallel or serially. To process in parallel means that two different functions are processing the same input at the same time and may correspond to different combinations of subsystems Additionally, a single system does not process a piece of data from beginning to end and then take in the next piece. As the first piece of data is making its way through the network the next piece is entering.
5. Opportunism: “We perform a task using whatever information is available, even if that information typically is not used in that context. For example.. .the parietal lobes of the brain may be specialized for guiding action; nevertheless, in some contexts the information used to guide action may also be used to distinguish one object from another. “6 Wet Mind researchers have applied their science to the areas of visual perception, visual cognition, reading, language, movement and memory. They have begun with computational analysis and model building. They have studied the underlying neurophysiology and the neuroanatomy. In particular this branch of science has looked at the overall behaviors of the organism. They have built up a comprehensive model of the brain and its functions. Each function has been modeled in neural networks and many combinations of networks have been assembled to study overall large Functions. The behaviors of the networks assembled have been put to various forms of behavioral testing. A considerable body of evidence has been collected by damaging these networks and seeing how they function in this state to the understanding of various forms of brain injury. Since one of the fastest expanding areas of the scope of practice in behavioral optometry is in the area of traumatic brain injury the work by the Wet Mind researchers is particularly enlightening.
My intent was to walk you through the results of the research done by this field. This is obviously beyond the scope of this type of presentation. I strongly advocate the reading of this book and for our profession to dialogue about the implications of the work in this field. I will conclude by showing a series of overheads and demonstrating some key aspects of the findings on this research.