These results demonstrate the feasibility of single-trial attentional evaluation using ANNs, but the paucity of data available for the current simulations necessarily renders this conclusion a cautious one. It is clear that some degree of success is attainable using ANNs, especially in the frequency domain; it is unclear what the limits of that success are and what performance expectations are reasonable. Compounding the lack of data was the low ceiling of the task; there were many fewer missed targets than correctly detected targets. In this preliminary investigation we did not design a task and collect data specifically to be used as input to these ANNs; rather, we borrowed data from another experiment. Having been through this dry run and established the feasibility of ANN analysis of attentional evoked potentials, we can in the future widen our focus to include the design of an experimental paradigm tailored specifically to this application. Such a paradigm will feature a more difficult discrimination between target and non-target stimuli, making careful attention essential to correct response and more nearly equalising the numbers of responses and missed targets in the data set.
The performance data on individual subjects are reported for the sake of completeness, but they are not to be trusted, because they are based on too few trials. The pervasive smattering of isolated, strong weights across the entire time epoch and the entire frequency spectrum suggests that much of the decision information was coming from chance correlations in the data set. The small size of the data set also may have impeded generalisation; it's possible that networks with more hidden units, trained on a sufficient amount of data, will be able to achieve higher levels of performance. Three-layer networks in some cases performed worse than two-layer designs on the same data set; in these instances the cost of implementing the extra constraints outweighed the benefit of a nonlinear solution.
The prominent difference between correctly classified target detections and misclassified target detections in autistic auditory averages may indicate a variable auditory response that occurs only in some trials. It is interesting to note that all the autistic subjects contributed to this variable response, and in none of their individual averages were the N270 and the P700 so well-defined as in this grand average of misclassified detections. Ironically, this set of misclassified responses with its extra peaks may represent a more normal response to target stimuli than the correctly classified responses. In other words, the network may have learnt to recognise the abnormal response because it was the more frequent one and therefore recognising it yielded a lower mean squared error than recognising the normal pattern. The variability of autistic auditory evoked potentials is known, and it has been observed that autistic brains do seem to have the capability to respond normally to auditory stimuli, though they manifest it only occasionally [Courchesne & al. 1984]. Due to the prominent P700 shift potential, one might guess that these misclassifications originated from trials in which a long time had elapsed since the previous cue to shift attention. Unfortunately, this hypothesis cannot be tested in the present study since information on specific inter-target intervals was not preserved in this analysis. In the future, it will be interesting to group trials by inter-target interval in a similar attentional-shift paradigm and to look for differences in response characteristics and in the classification strategies adopted by networks trained on these separate groups.
One particularly noteworthy result is that despite the attenuation and unpredictability of auditory P3 in the subjects with autism [Courchesne & al. 1984], two-layer networks in the time and frequency domains were able to reach performance levels of 81% and 85%, respectively, on the autistic auditory test set. P3 is the attentional marker most salient to the human eye, and it is interesting that the networks were able to attain this level of performance on a data set in which P3 was usually absent or severely attenuated. This suggests that either there is more information present in the autistic P3 that one might guess by looking at the averages, or that the networks were able to extract useful information from other features of the autistic evoked potential.
The likelihood that the autistic behavioural syndrome can arise from several distinct causes is widely acknowledged. Were this not admitted, it would be difficult to explain the tremendous amount of variance in the literature on autism. A current challenge in autism research is to identify and to characterise autistic sub-populations with different ætiologies. This problem can be approached at many levels--genetic, biochemical, anatomical, and behavioural. Analysis of the strategies developed by ANNs for the classification of autistic evoked potentials will introduce a new, physiological axis along which these groups may differentiate. It will be interesting to see what distinct patterns of electrical response might be picked out by ANNs, and how these electrophysiological differences correspond to differences at other levels of analysis.
The characterisation of sub-populations would rely on an examination of how networks accomplish the classification of responses. But the mere fact that such classification can be accomplished, by whatever means, might be of benefit to people who suffer from autism. If ANNs can perform better than other methods at extracting attentional information from the single-trial evoked potential, then the information transfer rate through a mental prosthesis may reach a practical level. The mental prosthesis then would be promoted from a laboratory curiosity to a real-world communication aid. Especially in the current atmosphere of suspicion regarding the communications of autistic people who use human facilitators, such a development would be very significant.