Are these two inequalities are equivalent?
Let's assume that $I_j \in \mathcal{J}$, where $\mathcal{J}$ is a set of images that are correctly classified and $p(I)$ is the output probability distribution of the used underlying model. Out of $\mathcal{J}$, we select $\hat{I}_{\!j^*}$ according to a well defined metric according to the inequality shown in (2). Given that $p(\hat{I}_{\!j^*} | \hat{y}= y) \leq p(\hat{I}_{\!j} | \hat{y}= y), $ where $y$ is the Ground Truth label. However, for $p(\hat{I}_{\!j^*} | \hat{y}= y) > p(\hat{I}_{\!j^*} | \hat{y}\neq y)$ and $p(I_j| \hat{y}= y) < p(I_j | \hat{y}\neq y)$.
How could we show that the first inequality shown in (1) is equivalent to the second inequality shown in (2)?
$$ \sum p(\hat{I}_{\!j^*})\log p(\hat{I}_{\!j^*}) \leq \sum p(\hat{I}_{\!j}) \log p(\hat{I}_{\!j}) \tag{1} $$
$$ \sum \log p(\hat{I}_{\!j^*}) \leq \sum \log p(\hat{I}_{\!j}) \tag{2} $$
A possible arugment that I believe might solve the problem is the following:
We can observe that the difference between both inequalities lies in the fact that (1) can be viewed as a scaled version of (2), where the probabilities are multiplied by their logarithms. Considering the monotonically increasing nature of the logarithmic function, we can roughly establish the equivalence of these two inequalities.
I have tested it numerically, and both inequalities are equivalent, but I cannot prove it mathematically.
- unanswered
- 295 views
- Pro Bono
Related Questions
- What is the Lebesgue density of $A$ and $B$ which answers a previous question?
- Rank, Range, Critical Values, Preimage, and Integral of Differential Forms
- applied probability
- Probability/Outcome
- Probability of choosing the bakery with the best bread
- Card riffle shuffling
- Statistics- Probability, Hypotheses , Standard Error
- A question in probability theory
Questions at this level should come with a bounty.