Linear model has worse results for dogs n cats?!

According to Kaggle, results for VGG + a linear model (with or without clipping to 0.05 and 0.98) has a worse score than the pure VGG results I submitted (with clipping). It’s not a huge difference (0.05 worse, or about 10 spots on the leaderboard), but it’s not the improvement I’d hoped for.

It was a useful exercise for learning the mechanics of Keras, etc., but I’m not sure about the bigger picture of why we would tack on a linear model this way. Leaving dogs and cats aside, which the original model with fine tuning was quite good at, I’m wondering if this technique is useful for other reason. Perhaps tacking on a the linear model to convert 1000 classes to 2 is more generally a handy technique? Especially if a model isn’t as confident (i.e. mostly 1s & 0s for dogs v. cats) to begin with?

What was the code you used in each case?

Also - try the same techniques on statefarm and let us know how you go with them, and what you think it implies…