A neural network model is presented which investigates the suggestion that being multilingual contributes to an individual's level of cognitive reserve. Two versions of this model were produced, one which learnt the names of input representations in a single language and another model which learnt input representations in two languages. The languages in both models were split further into two semantic categories. The relationship between the representations of both semantic categories in the first language of each of the two versions was investigated. Further manipulations came in the form of changing the sensitivity of the artificial neurons in the network and varying the hidden layer size with respect to variable levels of brain reserve. Findings were not immediately interpretable in terms of age-related decline in the absence of a behavioral measure. However, the variance in trajectories of category separation provide a cautionary tale against the interpretation of any measures gained at discreet intervals.
Recent studies of bilingual and multilingual individuals
have demonstrated some offsetting of normal cognitive
ageing
The contribution of language in offsetting age-related
cognitive deficits is one of a number of factors, known
collectively as the latent variable cognitive reserve
More recent studies of multilingualism and its association
with cognitive reserve have attempted to control for
education and intelligence.
To understand why bilingualism confers an advantage to
cognitive ageing, the cognitive mechanisms involved in
speaking more than one language must be unpacked. In
terms of nonverbal effects, these are wholly positive. Initial
findings in a study comparing English only speaking
Canadian children with their French- English speaking
counterparts on verbal and nonverbal tests found that the
bilingual children outperformed the monolinguals in almost
all aspects, especially the nonverbal intelligence tests
Whilst the cognitive advantages of bilingualism appear
well documented, the linguistic deficits associated with
having more than one language are equally well-researched.
For example, it is generally accepted that one of the
predominant negative effects of bilingualism is the
vocabulary size. This is generally smaller compared to
monolinguals for both languages spoken
The model in this study was a simple three layer,
feedforward back propagating neural network. Two versions
were produced, a monolingual version and a bilingual
version. The hidden layer was varied in size for this
investigation since it represented a view of passive reserve
Given the focus of study for the models was representation storage in the hidden layer rather than performance, a compromise between an artificial language and a realistic corpus was used for input. The inputs used in both models were patterns of 26 binary digits. The first 20 digits were randomised with the addition of a further six inputs. The first three of these represented a language tag. This was added to the experimental paradigm to guarantee separation of the two sets of pictures in the bilingual model since the rest of the input consisted of a random pattern. The final three binary digits of each input presentation related to the membership of semantic category A or B. 34 input patterns were used in the monolingual model and the monolingual input set was augmented with a further 34 patterns for the bilingual model, making 64 in total for the bilingual model.
The input ‘words’ used for the monolingual network were
taken from a dataset of English phonemes which had been
converted to a binary input set using a set of 19 features
Both networks were initially trained for 800 epochs. For comparison, test data was introduced to both monolingual and bilingual networks in the form of both categories of L1 only. 50 simulants were trained in this manner, the starting weights for each was seeded randomly from a uniform distribution of between 0 and 1. All of the following analyses represent mean scores. Training for both networks took around 200 epochs for the error to reach an asymptotic state. Overall, error settled at a slightly higher level in the bilingual network. This can be attributed to the increase in constraints in the bilingual network as it needed to accommodate the same amount of ‘pictures’ as the monolingual network but in both languages. Given that an asymptotic state was achieved around 200 epochs, it was decided that at 220 epochs the network was considered mature. It may be that there is a considerable ‘grace period’ during which the brain does not immediately decline after maturity. However, for the purposes of this study, aging can be said to begin upon reaching an asymptotic state, both for the purposes of analysis and interventions.
Using a methodology similar to
A C
B D
A C
B D
The main analysis within this study relied on the online calculation of the separation of categories A and B within L1. This was achieved through the calculation of a single centroid L1. This provided the distances to an overall mean. Further, distances to the representations within each category from the overall centroid were calculated. This provided a measure of within and overall variance upon which to calculate an F value. This was calculated at each epoch for each simulant over the four different hidden layer sizes.
The overall trend in categorical separation is driven by the feature differences between the two categories. Spreading of the categories can be seen progressing over the spectrum of hidden layer sizes for both models. However, the effect appears greater with the monolingual model. Conversely, clustering of representations within the hidden layer of the bilingual model is tighter. This was confirmed by cluster analysis of category A from L1 in both monolingual and bilingual networks carried out over the period of training (Figure 3.). This demonstrates that overall, higher hidden layer size networks showed the greatest spacing between representations with the two most dispersed categories belonging to the monolingual network. A 2*2 ANOVA carried out on the distances from the individual scores of the simulants at maturity with monolingual or bilingual as one factor and hidden layer as the other. The analysis demonstrated main effects for both hidden layer (F (3,392) = 1539, p<.001) and type of network (F (1,392) = 516, p<.001) together with an interaction between the two (F (3,392) = 1953, p<.001). Therefore, it appears that such an effect may be due to the differences in space cause by the additional constraints of second language storage offset only by a higher storage capacity. d tir o n e c m o r f s e c n a ti s d f o m u
S
In order to provide a more valid interpretation of
biological change over lifespan, the model integrated a
gradual decline in the slope of the log sigmoidal transfer
function. This manipulation reflects an age related
reduction in dopamine which in turn relates to cognitive
decline and a generally greater susceptibility to neural noise
as the network becomes less discerning (
Firstly, gain was set to decline gradually in steps of .0015 from the beginning of training. However, differentiation of either category could not be achieved. Therefore, a necessary and more valid representation of dopamine attenuation over lifespan was achieved by initiating the decline of gain after maturity, in this case, 220 epochs (Figure 4.) e u l a V F and bilingual models. Differences in the projections over time were investigated between the ten hidden layer networks for monolingual and bilingual networks. Firstly a bounded line graph was produced (Figure 5.) This demonstrated a greater variability in scores for the monolingual simulants. e u l a V F
Inline analysis of the separation items from both categories demonstrated an increasing separation of semantic information up until maturity for all groups. However, characteristics of the different projections differ in response to gain decline. Dedifferentiation appears almost immediately for ten and fifteen node monolingual models. However, for the other sizes of hidden layer for monolingual and all bilingual hidden layer sizes, dedifferentiation in the representational separation of categories for all hidden layer sizes for both monolingual
Multilevel analysis was used in which individual scores were used as the dependent variable with the epoch and as the first level predictor. The second level was grouped by whether the model was monolingual or bilingual. This was estimated as a random effect due to the differing projections between monolingual and bilingual.
Where Y represents the F-value at Epoch i for group j. A Likelihood ratio test demonstrated a significantly better fit with the inclusion of the random component ‘group’ (p<.001).
This study represents an initial attempt at exploring the way in which bilingualism influences how categories within languages are represented. The models presented in this paper represent the first attempt at modelling and analyzing multilingualism from a representational perspective. This study explored differences between monolingual and bilingual models, each with differing levels of hidden layer size, taken as a proxy of brain reserve capacity. Both models were trained over a number of epochs, representing a lifespan. Further, a change in the sigmoidal transfer function slope was included after asymptote. The results of this study provide commentary on the way in which picture representations separate according to input characteristics. Further, the results also provide an explanation of the behavioral observations of multilingual individuals.
The analysis carried out in this study suggests two main effects of bilingualism on semantic memory. Firstly, Categories within a single representational space in a bilingual speaker are sensitive to space. Equivalence on hidden layer size or brain reserve capacity leads to a greater contraction and overlap of representations that the monolingual equivalent. Secondly, gain change produces differing effects according to the constraining factors. Ten and fifteen node hidden layer monolingual networks declined in representational separation almost immediately. All other networks continued to separate to a degree before declining.
A trend toward poorer separation and greater clustering
together of the bilingual versions could lead to the observed
deficits in lexical access. The slower reaction times
observed in verbal fluency tasks
Given the executive nature of the positive contribution of multilingualism to cognitive reserve, it is difficult to relate the changes occurring in representational space to executive processes. However, what this study has indicated is that caution must be made when interpreting behavioral results gained from individuals over discreet periods of time. This is due to both the within group variance and the between category differences in projections over differing amounts of passive reserve and cognitive reserve, the former represented by differing amounts of hidden layer units and the latter represented by multilingualism. Specifically, a result of note is the steeper trajectory of decline near end of lifespan. This differed between mono and bilingual groups with a steeper decline observed in monolingual models. Further research is required to relate the representational spacing to behavioral outcomes. Further, it is important to note that the variability, especially within monolingual simulants, demonstrates the importance that must be placed on monitoring the progression of clinical outcomes for a single individual. The difference in projections demonstrate that behavioral outcomes may vary according to point in an individual’s lifespan that testing has occurred as well as the individual differences within a group. However, in order to fully relate performance to the changes in representational clustering demonstrated in this study, further research is required.
A novel analysis of the representational space was carried out on simple three layer networks portraying monolingual and bilingual speakers. The results demonstrate that some of the negative effects of multilingualism may be due to space constraints rather than relative underuse of multiple languages. Further, this research raises questions as to the efficacy of testing at discreet time points over multiple individuals where continuous assessment may provide a clearer picture of the contribution of multilingualism to offsetting cognitive decline.