IS4IS 2019 Berkeley

 

Semantic Information G Theory with Formulas for Falsification and Confirmation

Abstract: The semantic information G theory is a natural generalization of Shannon¡¯s information theory. Replacing yj in log(.) of Shannon¡¯s Mutual Information (MI) foviarmula with ¦Èj, a fuzzy set or a predictive model, we obtain the predictive MI formula. Using truth functions to produce likelihood functions, we have the sematic MI formula. We can also obtain this formula via improving Carnap and Bar-Hillel¡¯s semantic information formula Ij=log[1/T(yj)], where T(yj) is the logical probability of hypothesis yj. The improved formula is Iij=log[T(yj|xi)/T(yj)]=log[P(xi|¦Èj)/P(xi)], where xi is an instance, T(yj|xi)=T(¦Èj|xi) is the fuzzy truth value of proposition yj(xi), and T(yj) is the average of T(yj|x). Using a Gaussian function without coefficient as the truth function, we can find that logT(yj|xi) reflects deviation and testability. According to this formula, the larger the deviation is, the less information there is; the less the logical probability is, the larger the absolute value of information is; wrong hypotheses will convey negative information, and the information conveyed by a tautology or a contradiction is zero. Hence, this formula accords with Popper¡¯s thought about hypothesis-testing and falsification. To average Iij, we have the Generalizzed Kullback-Leibker (GKL) formula and the semantic MI formula. We can use the GKL formula and sampling distributions to optimize likelihood functions and truth functions for machine learning and induction. A hypothesis yj with a degree of belief b can be treated as the mixture of yj and a tautology with truth function bT(yj|x)+1-b. Using a sampling distribution to optimize b, we can obtain confirmation measure b*=[P(H|E)-P(H|E¡¯)]/max[P(H|E),P(H|E¡¯)]=[CL-CL¡¯]/max[CL,CL¡¯], where H=yj, E and E¡¯ are positive and negative instances respectively, CL is the confidence level, and CL¡¯=1-CL. The b* has HS symmetry suggested by Eells and Fitelson. It ensures that decreasing negative examples is more important than increasing positive examples and hence is compatible with Popper¡¯s falsification thought.

References: https://arxiv.org/abs/1809.01577 and https://arxiv.org/abs/1609.07827