next up previous contents
Next: 3.8 Conclusion Up: 3 The RF-LISSOM Model Previous: 3.6 Long-range inhibitory connections

Subsections

3.7 Environmental versus genetic factors in development

 

The RF-LISSOM model allows researchers to explore the relationship between environmental and genetic factors in development. The relative weights of these factors have been debated for hundreds of years (for review see Diamond 1974), but recent evidence is beginning to clarify how development actually occurs in the nervous system. This section will first sketch the very general constraints that apply to a developing organism and the tradeoffs involved in satisfying them. It will then examine how these tradeoffs might be optimized in the mammalian visual system in particular, and how they relate to RF-LISSOM. The intent is to show how a self-organizing model can explain aspects of both environmentally- and genetically-driven development. This helps to clarify the significance of the RF-LISSOM model, and to suggest the role that it can play in an explanation of the cortex.[*]

3.7.1 Evolutionary and developmental constraints

Every organism develops from a single cell. As such, it can make use of material within the cell, specifically the genome, and information in its environment, measured in some way. Different organisms appear to use these sources of information in very different ways, appropriate to their ecological niches and overall degree of complexity.

For instance, an animal's development could be specified in precise detail by its genome (Purves, 1988). This would be feasible if the important factors of an animal's environment remain essentially constant for millennia, and if the adult animal is not particularly complex. Each individual of the species would express exactly the same pattern, and evolutionary selection processes would ensure that the pattern is appropriate for the environment of the organism.

This developmental program appears to apply to the nervous system of the nematode worm C. elegans. Under ordinary circumstances, this organism has 302 neurons in the same configuration in every individual (Sulston and Horvitz, 1977). The advantage of such a scheme is that each individual will need only a fairly short developmental period, after which it will be fully prepared for life. A competing organism which requires a long maturation period would be at a disadvantage. While the competitor is still developing, the hard-wired organism will already have appropriate default behaviors even for circumstances that it has not previously encountered.

However, if the environment changes unpredictably over time scales shorter than those required for evolution to change the genome significantly, then a hard-wired organism will often encounter novel situations for which its behavior is inappropriate. An adaptive organism that can make use of clues found in the environment could be much more successful in those situations than a hard-wired ecological competitor.

It would be most efficient for the adaptive organism to extract all the relevant information from the environment at birth, so that it could subsequently devote its resources to other tasks. However, this would risk optimizing the organism for the particular circumstances surrounding that short time period. Optimizations made at that time might be wholly inappropriate for different situations encountered later, particularly since birth is an inherently anomalous event in an organism's lifetime. Thus it is desirable to integrate the information over as long a time as possible, to maximize the chances that the learned behavior is appropriate for the environment.

Yet the longer an organism takes to reach full competence, the more it is susceptible to accidents and predators. In the limit, the finite lifespan of an organism provides an upper bound upon how much environmental information can be integrated into the animal's repertoire. But since a single lifetime will only span a certain range of experiences, it cannot provide as much information as the millions of years that have presumably contributed to the formation of the genome.

If an organism became entirely adapted to the environment as it has encountered it, it would eventually encounter an important new event for which it was unprepared. In this case its hard-wired competitor would have the advantage. The genome of the hard-wired organism will presumably provide some appropriate behavior if the ``novel'' events actually recur quite often on an evolutionary timescale. So an organism cannot rely solely upon environmental cues, and to be fully suited to its environment it must (paradoxically) make use of some genetic information. The long-term information represented in the genome thus complements the direct environmental influences upon the organism after birth.

As discussed below, one might want the genetic influences to act as a constraint upon the amount of learning that can occur. This would prevent the organism from becoming too specifically adapted to its particular circumstances. The goal would be to help ensure that the organism remains capable of detecting unusual situations; e.g. those that are fatal but rare.

Thus there are a number of important tradeoffs between environmental and genetic control of development. For simple organisms with well-defined ecological roles, hard-wiring may suffice. However, more complex animals with more variable environments will optimally use some cues from the environment when developing. Fully adapting to the environment would not be entirely desirable, since it would amount to ignoring information only available in the genome.

Mammals, being quite complex and capable of handling a wide variety of different environments, appear to represent a particularly successful balance of these developmental and evolutionary constraints. Although significant genetic components are present, an extensive array of early environmental influences on visual system development have been documented in mammals (for review see Movshon and van Sluyters 1981). The following sections will argue that mammals make use of a simple technique that allows them to make use of both environmental and genetic information efficiently. This process will also be related to RF-LISSOM as a hypothesis for what the self-organization process represents.

3.7.2 Two-stage model for development

Since each individual begins from a single cell, yet some types of learning require multi-cellular sensory systems, genetic factors must determine the very first stages of development. Thus one simple way to combine genetic and environmental factors is for entirely gene-based development to progress to a certain point, e.g. birth or shortly thereafter, and then transfer developmental control to a learning mechanism (Blakemore and van Sluyters, 1975). This seems particularly appropriate in organisms such as mammals that have a long gestation period, during which their ability to measure the environment is severely limited. This approach would help minimize the time spent in learning, yet allow the environment to have some influence on the structures developed.

Once the developing systems are nearly complete, it would be desirable to limit their adaptability. Otherwise, eventually all traces of the initial genetic order would disappear (Jouvet, 1980). Allowing the organism to significantly adapt further would thus amount to ignoring valuable information from the genome (as described in the previous section).

In a computer model of sensory development such as RF-LISSOM, the hard-wired aspects would correspond to the initial state of the model, before any self-organization occurs. For instance, one could assume that a roughly retinotopic map on the LGN or visual cortex is formed through genetic means. Subsequent sharpening of the map could occur through visual experience, thus ensuring that the map is appropriate for the particular environment found.

However, this simple two-stage approach does not adequately account for all of the data on innate and environmental factors in mammals. Firstly, the genetic process cannot fully specify the initial phase of development in detail. There is simply not enough space available in the genome of a mammal (on the order of 105 genes) to specify every connection in its nervous system (as many as 1015 ; Kandel et al, 1991). Thus the genetic developmental mechanisms must express many highly repetitive structures or they must specify the structures only approximately (Purves, 1988; Shatz, 1996).

Secondly, even during embryonic stages, the developmental process appears quite flexible (Purves, 1988). Developing organisms adapt systematically to a number of rather drastic modifications, such as the implantation of a third eye in a frog (Reh and Constantine-Paton, 1985) or the removal of large areas of the brain of a primate (Goldman-Rakic, 1980). These adaptations presumably arise out of a repertoire of possible responses to damage or malfunction in the developing embryo. The embryonic adaptive mechanisms would be even more important for mammals than for simpler creatures, since greater complexity increases the likelihood of isolated malfunctions (Purves, 1988). Thus if each connection were hardwired, an explosive number of backup connections would need to be specified as well, or else some general adaptation procedure would be needed (Shatz, 1996). If there were such a general procedure, then the initial precise specification would be superfluous anyway.

Finally, evolution seems unlikely to have developed such a wastefully modular system, with two complex and non-overlapping systems to perform similar tasks. The first system would express genetic information according to a fixed plan. Since it must start from a single cell, it would need a set of bootstrapping mechanisms to construct the ``preset'' pattern determined by the genome. These mechanisms would have to be quite complex to account for the complex structures seen even at birth in mammals (Blakemore and van Sluyters, 1975). At some point determined by the genome, control would shift to the learning mechanism, and the cellular or intracellular structures used for bootstrapping would need to be deactivated or perhaps even dismantled. Otherwise, the system would not be able to accept subsequent cues from the environment. Instead of such a clean separation between genetic and environmental factors, it seems more likely that the initial mechanisms or structures are reused to some extent in later activity-dependent developmental processes (Shatz, 1996).

Cumulatively, the above constraints strongly discourage large-scale fully-specific hardwiring, yet somehow most organisms (including humans) develop with a large degree of structure already present at birth (Blakemore and van Sluyters, 1975). The next section explores an alternative, complementary hypothesis for combining genetic and environmental factors that addresses these considerations, and the following section will review evidence that it is actually occurring in several systems of mammals. This type of developmental strategy makes widespread use of activity-dependent self-organization, and so it helps to show how and why animals may implement the processes modeled in RF-LISSOM.

3.7.3 Overlapping model for development using pattern generation

True environmental information is only available to the visual system after birth (and often even later, depending on the time of eye-opening). However, coherent inputs arising at any point along the pathway from the receptor to the cortical areas could activate cortical areas. They could thus have effects similar to genuine visual experience, yet under genetic control (Constantine-Paton et al. 1990; Maffei and Galli-Resta 1990; Marks et al. 1995; Roffwarg et al. 1966; Shatz 1990, 1996). So once the basic, general connectivity and cellular mechanisms are in place, particularly the learning mechanisms modeled by RF-LISSOM, the system can self-organize based upon any input, not just those in the environment.

If the organism itself generates those signals with some form of internal pattern generator, the organism could direct its own development. Neural mechanisms to generate coherent patterns are well-documented in a number of different systems (Marder and Calabrese, 1996). Such a pattern generator would allow genetic factors to be expressed through the same mechanisms also used later when adapting to the visual environment (Jouvet, 1980).

As simulations with RF-LISSOM have demonstrated, even very simple patterns can drive the development of extraordinarily complex structures, such as the orientation map in V1 (described in chapter 4). For the orientation map, simple two-dimensional oriented Gaussian inputs, described by a single equation with six parameters, were presented to the general RF-LISSOM model. The model did not previously contain any representation of orientation. These inputs prompted the model to develop an organized set of thousands of orientation-specific feature-detectors, with many millions of specific connections between them. Thus quite detailed structures could be specified by the genome merely by specifying what type of pattern should be generated internally. Obviously, the nervous system might use any number of different mechanisms to generate appropriate patterns, not all of which would be as simple as the equation mentioned above, but in general much simpler mechanisms would suffice for the generation of training patterns than to specify the final structures.

Since the specification for the pattern generator could be encoded into a very small amount of genetic information (Jouvet, 1980), very different structures could be specified by changing only a small part of the genome. Random mutations in that portion of the genetic code would be likely to cause development of slightly different patterns, which might lead to quite different cortical structures. This would greatly facilitate the processes of evolution, since it increases the chance that a mutation is useful rather than merely debilitating.

To compare, imagine that the genome of a phylogenetically more primitive animal, such as a mouse, specified its nervous system in complete detail. The chance that it would eventually evolve into an animal with a more complex nervous system, such as a monkey, would be quite small. Any single change would be likely to damage the nervous system, yet an enormous number of coordinated changes would be needed to transform a fully-specific mouse genome into a fully-specific monkey genome (Purves, 1988). But even though much of the DNA is shared between mammalian species, an extraordinary variety of mammals exist, with quite different cortical organizations (Purves, 1988). Internally generated patterns may help to explain how evolution leapt over the gaps between successive species in phylogeny. This interpretation is supported by the observation that learning mechanisms appear to be highly conserved between species and brain areas in mammals (Kirkwood and Bear, 1994), while presumably pattern generators would vary significantly between different species.

The patterns of fur coloring on different individuals of the same species, e.g. domestic cats, might represent a graphic example of internal pattern generation in development. Swindale (1980) proposed a similar analogy between zebra stripes, fingerprint whorls, and ocular dominance stripes. Different kittens in the same litter often have very different fur patterns, some quite complex, yet each individual is genetically quite similar to the others. But purebred cats generally have the same overall fur patterns for every individual of that subspecies. The lack of variation in purebreds indicates that the patterns are controlled genetically, but the variety seen in mixed breeds suggests that the patterns are controlled by a small region of the genome, which can express quite different patterns with only small changes. This same type of pattern generation may be occurring in the mammalian nervous system, as explored in the next section.

3.7.4 Pattern-generated development in mammals

Retinal waves

  It is not presently known what role internal pattern generation plays in mammalian development. However, evidence is rapidly accumulating that such processes are occurring in at least two pathways leading to mammalian visual cortex: the developing retina and the brain stem. In the retina, the patterns take the form of intermittent spatially-coherent waves of activity across groups of ganglion cells (Meister et al., 1991; Sirosh, 1995; Wong et al., 1993). They appear to arise from as-yet-unknown events in networks of developing amacrine cells that provide input to the ganglion cells (Catsicas and Mobbs, 1995; Feller et al., 1996; Shatz, 1996). These waves may represent training inputs for the prenatally developing LGN and cortex, providing a simple explanation of how the system could be activity-dependent, yet already organized at birth.

The waves begin before photoreceptors have even developed (Maffei and Galli-Resta, 1990), so they do not result from visual input of any sort. They are presumably entirely unsynchronized between the two eyes, since they occur spontaneously and locally, with long pauses in between (Shatz, 1990). Blocking this activity in the retina prevents the segregation of the LGN into eye-specific layers before birth (Shatz, 1990, 1996), so the activity must be playing a role in the segregation. Blocking the activity in the retina, but substituting certain patterns of artificial electrical stimulation, results in dramatically different cellular properties than if other patterns are used (Mooney et al., 1993). This suggests that the patterns of electrical activity are the important feature of the waves (Shatz, 1990, 1996). As the waves subside gradually in strength and frequency due to developmental processes in the retina, the connectivity in the LGN stabilizes (Wong et al., 1993). At approximately the same time, visual input becomes functional, but it is not yet known whether visual input causes the waves to cease or whether they cease in anticipation of visual input. Manipulations of the visual environment have been shown to have dramatic effects on the organization of the visual cortex (Movshon and van Sluyters, 1981), the next step in the visual pathway. Thus internally-generated retinal waves appear to be instrumental in pre-visual development, at least of the LGN, and they appear to play a role similar to that which visual input plays for higher areas. It is not yet known what role the retinal waves play in the development of the cortex, but they may represent the prenatal portion of the activity seen by the cortex, which later receives visual input that is crucial for its development.

It has further been proposed that the spontaneous activity is responsible for maintaining topographic maps during the growth of the connecting pathways from the retina to the LGN (Bunt et al., 1979). That is, when the ganglion cells of the retina extend axons forming the optic nerve, they would keep a retinotopic order as a result of activity-dependent processes driven by these retinal waves. Throughout the developmental process, they would remain in this retinotopic order until they reached the LGN. This proposal has not been definitively established, but there is some evidence for the proposed mechanisms and results (Bunt et al., 1979). If it turns out to be well-supported, it would take still more of the burden off of the genetic factors in development. That is, since the retina is inherently ``retinotopic'', if the fibers remain in that organization throughout the journey to the cortex, no genetic specification of the retinotopic ordering is needed.

Finally, similar waves have been documented in early areas of other sensory systems, such as the auditory systems of birds (Lippe, 1994). Such activity may be a general feature of the earliest sensory areas in more complex organisms such as birds and mammals, allowing genetic factors to be take effect via the same mechanisms which later incorporate environmental influences (Shatz, 1990). This would represent a developmental mechanism which is quite efficient and effective on a number of different levels, as described above.

PGO waves

  Internal pattern generation may continue at somewhat reduced levels, even after development. For instance, the generation of appropriate training patterns has been proposed to be one function of the sleep stage known as rapid eye movement sleep (REM; Jouvet 1980; Roffwarg et al. 1966). Infants exhibit very large durations of REM sleep during the ages when their nervous systems are the most highly plastic, and the amounts of REM sleep and plasticity decrease similarly with age (Roffwarg et al., 1966). Furthermore, birds and mammals, which are much more adaptable to different environments than other animals, are the only organisms known to exhibit REM sleep in the adult (Jouvet, 1980).

During and just before REM sleep, internally generated phasic waves called ponto-geniculo-occipital (PGO) waves can be measured in the LGN, V1, and many other cortical areas (Jouvet, 1980; Steriade et al., 1989). These apparently genetically-programmed waves consist of large, slow increases in field potential when measured with an EEG (electroencephalogram). They appear to cause specific eye movements which have been statistically correlated with reports of the direction of gaze in dream imagery (Jouvet, 1980). PGO waves originate in the pons of the brain stem (hence ponto-) and travel via direct pathways to the lateral geniculate nucleus (hence -geniculo) and to visual cortex (in the occipital lobe, hence -occipital; Steriade et al. 1989). They appear to be relayed from the LGN and visual cortex to many other areas of the cortex (Jouvet, 1980).

Jouvet (1980) has proposed that these waves help direct the course of brain maturation in early life, and specifically that they allow genetic differences among individuals to be expressed. Tentative support for this hypothesis was obtained by Marks et al. (1995), who found that depriving kittens of REM sleep during the critical period significantly enhanced the effects of visual experience during that time. The heightened effects of experience were interpreted as a weakening of genetic control over development.

Ordinarily, blocking signals from one eye of a kitten for even a few days during the critical period for visual development (4-6 weeks after birth) causes dramatic strengthening of connections to visual cortex from the non-deprived eye, and a loss of connections from the deprived eye (Movshon and van Sluyters, 1981). With shorter deprivation times, the effects are less pronounced. If both eyes are deprived of input, no detectable loss or strengthening occurs (Movshon and van Sluyters, 1981). Thus a monocular deprivation paradigm offers an opportunity to test the role of environmental and genetic factors in development, both of which appear to be operating in this system.

Marks et al. (1995) blocked signals from one eye of kittens for a portion of the critical period. It was found that blocking the input had a much larger effect on animals deprived of REM sleep than for control animals which had normal REM sleep. Similarly heightened sensitivity to deprivation was found when the PGO waves were blocked directly by lesions in the pons. Thus some mechanism in which the PGO waves participate appears to limit the amount of plasticity in this system in the normal individual.

Unlike the retinal waves discussed in the previous section, PGO waves are correlated between the two eyes. The patterns sent to each of the eye-specific layers of the LGN are generated separately, but match quite closely (Jouvet, 1980). The correlation presumably represents the pattern of similar, but not identical, eye movement directions for the two eyes. This type of correlated input would ordinarily counteract the imbalance of inputs from the deprived and non-deprived eyes. Thus the monocular deprivation protocol should have had greater effect in the absence of PGO waves, which was in fact seen (Marks et al., 1995).

Jouvet (1980) speculates that PGO waves are a very general signal that activates local circuits in each brain area. Each local circuit would then generate training inputs appropriate for its area. He further proposes candidate substrates for such circuitry, which are beyond the scope of this discussion. However, if Jouvet's proposal is correct, then to the extent that the pattern evoked by PGO waves differs from the typical response to the environment, REM sleep would serve to ensure that each brain area develops and is maintained in readiness for its genetically-determined function. The genetically-determined structure would be fine-tuned based on information from the environment, but it would persist regardless of how much learning occurs. This would help ensure that the structure and function of each area of the brain would be suitable for the species of the organism, and thus for its ecological niche. It would also help differentiate different brain areas intended for different tasks yet which receive similar inputs from the environment; without such differentiation or some form of competition the areas would eventually become identical as they adapt to visual input.

As argued above, it is not desirable for an organism to completely adapt to the environment it experiences. The amount of time available for adaptation is very small compared to the time over which genetic information has been compiled, and thus not all the relevant information is available from the environment. Thus the PGO waves may also serve to limit such unwarranted adaptation. Since adults exhibit significant amounts of REM sleep as well, these pattern generators may also be operating in the developed animal (Jouvet, 1980). They may help maintain a balance between specific visual correlations learned during the day, such as a preponderance of certain orientations of lines, and circuitry capable of handling a wide variety of processing tasks, such as detecting all possible lines. Roffwarg et al. (1966) and Steriade et al. (1989) speculate that these generated patterns constitute some of the vivid imagery experienced during REM sleep.

Jouvet (1980) further speculates that the function of REM sleep is to ``program'' the nervous system with genetically encoded behaviors and capabilities. In preliminary experiments, he tested the effect of REM sleep deprivation on two strains of genotypically similar laboratory mice. These two strains ordinarily exhibit a small number of well-defined differences in maze-learning behavior even when raised in identical environments. When deprived of REM sleep, the difference in behavior decreased, i.e. the mice became even more similar. Although the experiment was not conclusive, it suggests that the REM sleep deprivation inhibited the expression of genetic differences between the two strains. In general, Jouvet hypothesizes that individuals would exhibit far less variation if deprived of the effects of REM sleep, since their behavior would be determined primarily by the environment. Thus ``genetic programming'' during REM sleep may help to ensure that different species, individuals of the same species, and even different areas of a single brain retain diversity, thus increasing the likelihood that some member of the group will be appropriate for a given task.

As a cautionary note, the REM sleep effects are not nearly as well characterized as those in the retina, primarily because sleep deprivation has a number of side effects. For instance, both behavioral and pharmacological methods of REM deprivation cause significant stress, and the pharmacological methods involve drugs that are not particularly specific in their actions (Jouvet, 1980). This makes REM sleep results difficult to interpret. The PGO waves themselves are well-documented, however, and since they occur in many different areas of the brain, they could represent a general feature of mammalian neural systems. They may be a part of the mechanism by which genetic and environmental influences are combined by activity-dependent learning processes.

Although research into the effects of pattern generation on the developing brain is just beginning, it is already clear that the distinction between ``genetic'' and ``environmental'' origins of brain structures is blurring. Many alternatives in between can be determined by the ratio of pre-programmed ``experience'' to actual sensory experience. The use of pattern generators to direct experience-dependent learning processes represents an effective solution to the general constraints faced by a complex organism with a limited genome and a limited time for development.

3.7.5 Pattern-generated development in self-organizing models

Self-organizing models like RF-LISSOM allow hypotheses about the contribution of genetic and environmental factors to be tested in detail. The training inputs for self-organizing models can represent either internally generated patterns or environmentally realistic stimuli, or any combination thereof. RF-LISSOM also incorporates hard-wired genetic information, such as the basic learning and activity computation algorithms, which are assumed to be essentially the same between different individuals.

The initial connectivity parameters, such as the extent of excitatory and inhibitory connections or the amount of initial order, can be set by the researcher to represent different possible starting states. The RF-LISSOM research so far has tried to begin from as much initial randomness as is possible for self-organization. This shows how RF-LISSOM can explain the most difficult cases. Adding further initial order, representing some innate bias towards the appropriate structure (as often found in biological systems), just makes the learning process easier.

Similarly, simulations are run using the simplest training inputs that will account for the structures seen in the cortex. These inputs can then be compared with known internal pattern generators, and with the typical features of the visual environment. Finding that the required inputs match with those available from either of those two sources helps confirm that the model is appropriate. Using such techniques, the RF-LISSOM model can help test hypotheses about the relative importance of environmental and genetic factors in development.


next up previous contents
Next: 3.8 Conclusion Up: 3 The RF-LISSOM Model Previous: 3.6 Long-range inhibitory connections
James A. Bednar
9/19/1997