Another cold and windy morning, but unhappy on top

 So the morning was somewhat more sunny but I just couldn't bring myself up from some depths of melancholy so there's that.

Just checked gradcafe and my email, nothing interesting anywhere.

I should be working but as usual I don't feel like it. I did study some theory of empirical risk minimization.

I have been feeling like crying since the morning, but it isn't going to help.

Someone advised me to just stop thinking about things, and I'm trying that.

I wish I could get a coffee from crazy mocha. Not gonna happen right now.

There should be a simple matrix for describing how one is feeling at any given moment.

I guess I'm thinking about robustness of deep neural networks all the time these days. Yesterday I read the original paper by Szegedy, and it was a nice piece of work. They hypothesize that these are due to statistical nuances of the training dataset. I find the whole thing fascinating from the point of view of being clear about what a trained neural network can and can't do. Also I think they represent generalization errors, as in the network learns to associate something random with a particular label and carries that error forward in its lifetime of classifications. What does the decision surface of a neural network look like is another question we should be asking ourselves. For images, since random and small perturbations cause misclassifications, it seems that the decision surface is extremely jagged.

Another question we should be asking ourselves is whether the increase in performance due to, say, a larger network, comes at the cost of generalization. Sure, you can use validation error for early stopping, but some generalization errors are likely to reward you on the validation set too. I am not very familiar with vision problems in general, but I do think that large networks are extremely extremely overparameterized. One thing I noted is that Sazagady mentioned that adversarial examples transfer to models that have been trained on a different subset of training data, but the success there is very small and further references ignore this fact.

Overall, I am in favour of developing global techniques for interpreting a model, but I don't know what the current state is in computer vision. After training, the model should provide a summary of what it has learned to associate with each class, and what is anti-correlated. NLP problems are surprisingly simpler, despite what we claim in our work, but my lack of access to a GPU doesn't let me work on vision. 

Another interesting line of thought stems from the fact that data has some real world footprint, and we should be aware of it in our practices. For images, I am very keen on observing the effect of perturbing edges on network output, but there doesn't seem to be any correlation. Ugh!

For a fixed input CNN, theoretically we can test all different permutations of input on the model and observe its output. For discrete input/discrete output systems, this input/output map would be a valid summary. Pixel values aren't really continuous and there is a theoretically finite number of inputs you can provide. Sampling from this distribution is another problem however.