False feature hypotheses

FIGURE 8.4 ABIS. Pure extraction of edge elements (i.e. pixels with high gradients of intensity) yields correct feature hypotheses, but also many positive false feature hypotheses due to hairs, polls, etc. 

Note that performing multiple hypothesis tests (e.g., Student s t-test) may inflate the false positive rate. That is, the number of genes detected as active by chance alone will increase with the number of genes tested. For example, a microarray with 7000 features would require at least 7000 hypothesis tests per treatment comparison. Several methods have been developed to control the false positive rate, such as the conservative Bonferroni correction and the FDR control method (49). 

Frequentist statistics (Also frequently and classically referred to as classical statistics.) An approach to statistics commonly encountered in the analysis of clinical trials and having the following features. (1) An interpretation of probabilities is made which relates them to the long-run relative frequency of events in a series of (hypothetical) trials is made. (2) It is denied that hypotheses or parameters can have a probability the former are either true or false and the latter are either equal to some value or not. (3) A decision to accept or reject a hypothesis (or presumed parameter value) is made indirectly using the probability of the evidence given the hypothesis (or presumed parameter value) rather than vice versa. (4) The probability of more extreme evidence must also be taken into account. (5) The experimenter s intentions in designing the trial have a bearing on the interpretation of the results. 

In general this model is quite effective as it solves the feature explosion problem by positing a modular approach. One significant weakness however is that the elasticity hypothesis is demonstrably false. If it was true then we would expect the z-scores for all the phones in a syllable to be the same, but this is hardly ever the case depending on context, position and other features the z-scores for phones across a syllable vary widely. This is a problem with only the second component in the model and a more sophisticated model of syllable/phone duration interaction could solve this. In fact, there is no reason why a second neural network could not be used for this problem. 

---