> "72% and 89% are as close as 75% and 50% in terms of being double the likelihood for a Republican win."
In a postscript (P.S.) Andrew Gelman clarified that the forecast distributions are close and not the probabilities of winning [0].
> "I’m not saying that 72% and 89% are close in probabilistic terms. I agree that there’s a big difference between 8-1 odds and 3-1 odds. What I’m saying is that the large difference in these probabilities arises from small differences in the forecast distributions. Changing a forecast from 0.54 +/- 0.02 to 0.53 +/- 0.025 is highly consequential to the win probability, but as uncertainty distributions they are similar and they are hard to distinguish directly. You can get much different headline probabilities from similar underlying distributions."
TLDR: The use of texture is inherent to the ImageNet dataset and not to deep learning / ConvNet. Training on less-textured versions of ImageNet drives the ConvNet to focus more on shape.
Sure, but just because you can coerce an algorithm into doing something different does not mean the algorithm itself does not have fundamental tendencies.
