Page 155 - Artificial Intelligence in the Age of Neural Networks and Brain Computing
P. 155

144    CHAPTER 7 Pitfalls and Opportunities in the Development of AI Systems




                            The data aren’t relevant to our problem. Doubtless there is a lot of information in
                         our data, but does any of it pertain to our task? We are trying to use information in our
                         data to separate two classes of cases. If that information is lacking, then no matter
                         what kind of CI observer we develop it just won’t perform well. We lack discrimina-
                         bility. Perhaps we are using low resolution images to try to solve a high resolution
                         task. Perhaps we are using imaging data for prostate cancer detection when the
                         amateur neighborhood proctologist (Fig. 7.5) is right in suggesting that olfactory
                         data would lead to superior performance.
                            Our data aren’t internally consistent. Large datasets are often composed of
                         “meta data” consisting of many chunks of data from different sources. For example,
                         large-scale NIH clinical trials are frequently criticized for being based on an
                         agglomeration of data from different institutions using different medical equipment
                         and following different clinical protocols. Or, the data may simply be out of date.
                         Sensors are continually being improved and made more sensitive and less noisy
                         with better spatial, spectral, and/or temporal resolution.
                            Our dataset isn’t representative of the universe on which our CI observer will be
                         used. The data are fine but not directly relevant to our task. Again in medical terms,
                         data from a diagnostic study (performed on sick people) may not be applicable to
                         a device used in the screening of an asymptomatic population. Or in the CGI world,
                         it is composed of views of normal human landscapes, but it will be used for



























                         FIGURE 7.5
                         Human modesty has largely robbed modern medicine of the use of the olfactory sense.
                         We generally view it as offensive to sniff or be sniffed. Dobie, however, does not feel that
                         constraint. There have been numerous news reports touting canine abilities in cancer
                         detection [5], and there is a very active field of development of artificial noses [6].
                                                                          Image Garosha j Dreamstime.
   150   151   152   153   154   155   156   157   158   159   160