Let us assume that the element oxygen has been discovered and that the method by which atoms are distinguished from one another is to count their protons. Still, puzzling observations, which cannot be predicted by relying on proton numbers alone, will be encountered as soon as ozone (O3), molecular oxygen (O2) and free oxygen radicals are in our air sample. To get the most predictive power and highest F-score, investigators tend to try to optimize prediction by trying to find a method to predict the most common outcome. Accordingly, in this oxygen example, researchers tend to develop an algorithm, which is best to predict bimolecular oxygen (O2) because it is most abundant among all oxygen molecules (i.e. ozone (O3), bimolecular oxygen (O2) and the negatively charged oxygen radical (O-)). The error rate under the assumption that there is no difference between oxygen molecules would be equal to ozone/bimolecular oxygen + oxygen ions/molecular oxygen. In order to distinguish between these different kinds of oxygen the electron/proton ratio, the different charge distribution on the molecular surface, molecular weight, molecular volume, the arrangements of chemical bonds, and the position of the oxygen atoms relative to one another within the same molecule, could be added to our training data in order to distinguish between the different oxygen species. But let us assume that we are still naïve and cannot measure the needed features yet, how could we go about discovering the missing/hidden features? In general, varying the features of the input training data for training a supervised machine algorithm, the learning steps, the inert environment and the methods of measurement must be selected based on intuition due to lack of any better alternatives. For AI to correctly determine the overall electrical charge of an oxygen molecule, AI needs the number of protons and electrons as input data. Unfortunately, if the instruments for detecting protons, electrons and neutrons are lacking, we can see the effect of the still hidden factor, i.e. electron/proton ratio, on the overall molecular charge but its reason still remains a hidden mystery. In this case, investing time to discover electrons, neutrons and protons, is much wiser than trying to tweak the parameters after the error rate has reached its asymptote, because even if this improves prediction, there is a big risk of over-fitting, because AI is basing its decisions on features, which actually have no effect on the overall molecular charge. But instead of using the electron/proton ratio as input features, the molecular size of the different oxygen species, would also work for training our AI-molecular charge predictor. Electron/proton ratio (a simple fraction) and molecular size (a volume measured in cubic nanometers) are different dimensions; yet both of them can express the same event, i.e. electric charge. Therefore, both could be used to train AI on predicting the molecular charge correctly. If, like in the example above, the in reality observed outcome can be perfectly predicted in at least two different dimensions, then it is reasonable to believe that all hidden factors have been discovered. The relationship between electron/proton ratio and molecular volume is about the same as between transcription factor binding sites (TFBS) and the trajectories of gene expression time series plots
To support my hypothesis that we are still many concepts away from understanding and manipulating aging, I give the example of the mesentery. It took humanity until early 2017 to discover its function and group it together differently according to the new discoveries (citation needed).
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence