Introduction
Since the dawn of mankind, people have been mystified with the aging process, attempting to find a way to avoid its inescapable reality. Throughout the ages, myths have centred around conquering this eternal plague, from the obsession with the elusive fountain of youth, immortality drugs, and the host of reincarnation beliefs found in a variety of cultures. These persist to the present day: though the general life expectancy in North America has more than doubled in the last two centuries, an obsession with age can easily be seen. Advertisements are often heard in the media that “[this wrinkle remover] will make your face look younger”. As well, many anti-aging drugs are on the market, including ones that quench the harmful free-radical process, which is said to be a primary cause in the brain’s deterioration. A wide variety of nootropic drugs can be found with a brief Internet search; these are dubious drugs that potentially increase the functioning of the brain, slowing the advent of aging. Yet one avenue of potential hope has been largely disregarded by the press: the potential role that computers can play in both aging and disease research via the use of cognitive simulations.
Aging
Many people believe that as one gets older, it is inevitable that one’s memory and cognitive ability will diminish in capacity and in quality. Yet, this is untrue. This is just one of many popular misconceptions that people have of the aging process, arising out of cross-sectional studies performed in the early days of psychological testing. These were studies that compared the general cognitive functioning of people in different age groups at the same time. They were often dependent on specific knowledge and did not account for such factors as the differences in teaching processes in different generations.
More recent longitudinal psychological tests have shown quite different results. These are tests that compare results of a single individual at different ages, over an extended time span lasting years or decades. It has been determined, that, for a majority of the elderly, some cognitive tasks are performed better as they get older, demonstrating that learning occurs throughout a person’s life. These include improvements in vocabulary and the sequencing of pictures. However, there were noticeable declines in their overall speed at processes, and other specific tests. These included a test where they are asked to draw a person as accurately as possible: their quality of picture formation decreased, as did their ability to complete sentences with appropriate words and clauses. In the majority of cases, it was found that the only difference in performance was speed, and when they were given enough time, they would demonstrate similar performance as they became older. Reasoning was found to be one of the most stable cognitive functions throughout the aging process. (Cohen, p. 25)
As a counterpoint, it has been determined that a distinction must be made between fluid intelligence and crystallized intelligence in regards to aging. Fluid intelligence can be thought of as acquiring new concepts and adapting to new situations, whereas crystallized intelligence refers to having previously learned skills. It has been found that fluid intelligence is affected much more severely with age than crystallized intelligence. (Parks, p. 7) Hence the adage, “You can’t teach an old dog new tricks.”
Internally, even in a healthy elderly person, it can be seen that, in general, both the weight and volume of the brain gradually diminish by a slight amount after the age of fifty, with some localized atrophying in parts of the hippocampus and amygdala, which are known to be responsible for emotion, yet it is unknown what relation this has to any external symptoms seen in the aged. Widely differing results have been seen in different studies on the amount of neuronal loss in the normal aging process, but in general, the loss seems to be minimal, and the amount of neuronal shrinkage has often not been fully taken into account in studies where greater loss has been reported. (Esiri, 1994)
Alzheimer’s Disease and Dementia
General memory loss and dementia do affect a significant portion of older adults, with the causes for many of these problems not being fully understood. For example, Alzheimer’s disease, the most well known and most commonly occurring form of dementia, is clinically diagnosed by a patient showing a lack of symptoms of other dementias, and can only be diagnosed after all other possible causes have been overruled or posthumously via autopsy. Without a detailed understanding of possible ways memory loss can occur, it will be impossible to treat Alzheimer’s any sooner than it is currently. It can be hypothesized that with simplified computer simulations of aging and memory loss, it would be possible to more locally determine the potential causes for Alzheimer’s disease and other dementias, and could therefore help lead to a better understanding of this problem and what is occurring in the mind as these diseases progress.
Externally, there are many symptoms of dementia and countries have differing criteria for classifying it. By the International Classification of Diseases, tenth revision, the four diagnostic criteria that must be fulfilled include (Henderson & Sartorius, 1994) :-
1) A decline in both memory and intellect, with mild, moderate, and severe levels of impairment in each
2) Absence of clouding of consciousness
3) Deterioration in emotional control, social behaviour or motivation
4) Criterion 1) must have been present for at least 6 months.
Mere forgetfulness and minor memory omissions have been noted as often being benign, and not leading to dementia. Yet dementia can be triggered by emotional stress, strokes, diabetes, and other diseases, both related and unrelated to cognitive dysfunction.
General symptoms of dementia include: memory loss, dysnomia and visuospatial deficits, with impaired intellectual functioning. Dysnomia is a condition where it is difficult to remember names or recalling appropriate words to use in a given context. Visuospatial deficits include the inability to mentally manipulate objects and perceive differences in object orientation. A person experiencing a mild or even moderate case of dementia will generally not have all these symptoms, but will usually show some kind of memory deficit. Episodic memory is also severely affected. (Parks et al, p. 11)
Internally,
through neurological studies, a limited understanding of the factors
responsible for the cognitive disabilities can be determined, but not their
specific causes. Neurons that are
affected by dementias such as Alzheimer’s disease contain an increase of
neurofibrillary tangles, neuropil threads and senile plaques. (Esiri & Morris, p. 7) Neurofibrillary tangles and neuropil threads
are abnormal insoluble formations that are allowed to build up in the neurons
which do not realize their harmful nature.
Senile plaques reside outside neurons and reduce the ability for neurons
to communicate with each other. It is
unknown exactly why these symptoms occur or what produces them, whether it is a
cell’s reaction to the disease or the disease itself.
Cognitive Model
To simulate the associated learning and recall that is performed by the brain, a feedforward network with backpropagation was used as the base neural network. In feedforward networks, information is only allowed to flow in one direction on multiple levels of neurons. Input is set specifically, and then it is allowed to pass into the hidden layers of the neural network, until the output layer is reached and can be examined for the output of the neural network. Error backpropagation is used to correct errors in the network and to stabilize it: this is done by calculating the difference between what the output should have been and the output actually produced, and then this error is recursively minimized by changing the weights of the associated hidden layer inputs until the first input layer is reached. This is done after every response by the network, while it is being trained. When it is time to actually test the network, error backpropagation is not further used and the outputted results of the network are taken explicitly. (Muller et al, p. 53) Because feedforward networks use a distributed representation of information, it was felt that this base model would correspond reasonably well to a human brain, as determined by current theories.
To simulate the advent of aging and dementia on the brain, different operations were performed on the network. These included: randomly deleting links between neurons and deleting neurons in their entirety.
Randomly deleting links between neurons is meant to simulate the destruction of neural links by the external senile plaques which are a major correlate of the appearance of Alzheimer’s disease. This is an appropriate approximation because the plaques reduce the ability of the neuronal links to work appropriately; they are unable to send signals effectively.
Systematically lowering the weights of links between neurons could also have been done: this would simulate the reduction in effectiveness of neuronal links and the greater activation needed in neurons to fire when they have aged or when neurofibrillary tangles and neuropil threads are affecting them. Reducing them a small amount simulates the aging process, and increasing the proportion of links that have lost ability simulates a person getting significantly older, or developing Alzheimer’s disease or another form of dementia.
Deleting neurons in their entirety is simulating the effect of the atrophying on the neurons in later stages of Alzheimer’s disease. As described earlier, particular neuronal clusters in the mind atrophy, making them useless and unable to send any signals.
If a localized network had been used, the effects would have been obvious because of the inherent structure of the model. There is no associated distinct model for the fluid and crystallized learning in this case because the network is pre-constructed. As links between different items were deleted while learning was in progress, those items would simply lose their associations and the network would be biased towards the results that would have occurred without the associations ever having been existed.
Critical Assessment
There are many problems with the cognitive model in question, both on macroscopic and microscopic levels. In general, the model does not factor in the representation of chemical effects in the brain due to aging or dementia, it does not factor in the underlying emotional effects of the person experience these maladies, and it does not provide much in the way of counterbalancing effects, which have been demonstrated by humans. The model also assumes that input and output neurons do not fail. There is also no idea of neuronal cluster interaction in the model and the effect this would have in the brain, when different clusters are aged or fall to dementia. As well, the entire representation of the brain by a neural network can be called into question, and also the lack of any effect similar to backpropagation being seen in the mind.
Deleting the dendritic links in a computerized neural network demonstrates the electronic effects of again and dementia, but fails to account for the myriad chemical effects that are occurring alongside the electronic. It is unknown what chemical affects are being caused by the variety of symptoms of Alzheimer’s disease or aging, for that matter. For example, substances that remove free radicals, which are thought to be catalysts of the aging process, inhibit the creation of senile plaques. (Esiri & Morris, p. 107)
Not only are chemical effects not accounted for, but neither are emotional effects. When a person feels the onset of age, there is little sign of additional negative feelings that is not seen in young adults. Each time period comes with its own amount of worry and strain, and there are no general emotional differences. It has been seen that in old age the fear of death is surprisingly uncommon. (Cohen, p. 35) But, with dementia, as with other disorders, an elderly person can find themselves having repeated depressing thoughts about death or life in general. This can have further untoward effects on the mental state of the person, leading to a cyclical effect and further health problems. A computer simulation fails to register the effect emotions play on the cognition of a person. Emotions have been thought of as being an overall diagnosis of the mental state of a person, and as such, they could help counterbalance the person’s dementia as it arises. The model does not take this into account.
As well as the macroscopic effect of emotions, microscopic counterbalancing effects come into play. If particular neurons are being destroyed, other neurons could perhaps take over their functionality. It is quite possible that there are immense backup systems in the brain: this has been seen in the ability of people to recuperate after loss or damage of one of the major sections of the brain. This can be seen to a limited extent in the model, but not to the degree the brain demonstrates: because a small, toy model is being used, the degree that counterbalancing effects can be displayed is limiting.
Another problem that is not accounted for by the cognitive model is that of the failure of input and output neurons. In the model, it was determined that these were immune to the aging and dementia process because if they disappeared, then the resultant effects would be obvious: there would be no resulting input or output, depending on which neuron was removed. Because the brain has such a large number of neurons, operating in parallel, a single neuron can be hypothesized as not providing a large amount of functionality, whereas in the model, a single neuron, especially an input or output neuron, has a much more significant effect. This is indeed another limitation of having such a small model. The interactions between different neuronal clusters in the mind could also help alleviate the problem of the failure of input and output neurons because there could be such things as error correction mechanisms that the brain uses when encountering erroneous input. When data is being transferred over the Internet, for example, there will be checksum bits that are used to ensure that the information being transferred is correct, despite loss of packets of information and miscommunication. The brain must have some kind of self-checking mechanism for input and output; this can be seen in its ability to correctly interpret partially covered up symbols, for example. Once again, the model fails to account for this kind of error correction.
The model is reasonable on many other accounts, though. This type of aging model will work with any kind of distributed network, as it requires solely deleting and removing the neurons in a gradual fashion. It also scales up to much larger and more complicated examples because of the general nature of the model: as long as there are links and the internal representation is a distributed one, the model would continues to be an effective simulation. It accurately simulates the electronic effects of neuronal loss in the mind and corresponds to the results found in people, as far as any standard neural network. It accurately shows the deleterious effects that dementia has on fluid learning and the ability to recall information.
It can be questioned whether any model of cognition using backpropagation is an appropriate one. There has been no display of backpropagation being used in the mind itself and many tasks, such as language comprehension, are learned without the use of direct, consistent feedback or any error-correction mechanism. The analysis of the underlying functionality of large neural networks is purportedly difficult and the need for randomness when initially generating the model and for determining an appropriate number of units to be used in a simulation also demonstrates imprecision, and thus flaws in the model. Any model that trains itself will be subject to these flaws. (Coltheart, 1995)
Conclusion
Removing both the links between neurons and neurons themselves in a neural network using backpropogation provides an approximation to the memory impairments and learning deficits that are seen in the onset of aging and dementias such as Alzheimer’s disease. Though this model only takes into account the electronic nature of the neurons, without associating itself with untoward chemical or electronic effects, it still provides an effective simulation that can be generalized to any trained or distributed neural network.
References
Cohen, Gene D. The Brain and Human Aging. New York: Springer. 1988.
Coltheart, Max. Connectionist Modelling and Cognitive Psychology. Noetica: A Cognitive Science Forum. Domain: “http://www.cs.indiana.edu/Noetica/”
Website: “http://www.cs.indiana.edu/Noetica/OpenForumIssue1/Coltheart.html”. 1995.
Esiri, Margaret M. Dementia and normal aging: neuropathology. In: Hupert, Brayne & O’Connor (Eds.), Dementia and Normal Aging. Cambridge University Press. 1994.
Esiri, Margaret M. & Morris, James H. (Eds.) The Neuropathology of Dementia. Cambridge University Press. 1997.
Henderson, A. S. & Sartorius, N. International Criteria and Differential Diagnosis. In: Hupert, Brayne & O’Connor (Eds.), Dementia and Normal Aging. Cambridge University Press. 1994.
Muller, B., Reinhardt, J., & Strickland M. T. Neural Networks: An Introduction. Berlin: Springer. 1995.
Parks, Randolph W. et al. (Eds.) Neuropsychology of Alzheimer’s Disease and Other Dementias. New York: Oxford University Press. 1993.