3 ms·
Question: do these blind spots of neural nets persist as the data set is expanded? I understand that if you have a fixed training data set of, say, 10 million
by michaelochurch 12y ago
Question: do these blind spots of neural nets persist as the data set is expanded?
I understand that if you have a fixed training data set of, say, 10 million images, there are probably going to be adversarial examples that will be specific to a net trained on that set. (These may be the same artifacts that are picked up by overfitting, or they could be local underfitting artifacts due to early stopping.) But are they specific to the data set?
In other words, are the same adversarial examples going to "break" a neural net trained on a separate (but identically generated) set of 10 million images? Or is this something that can be smoothed away with, say, ensemble methods and multiple data sets?
- pfortuny 12y agoI guess this boils down to something similar to the approximation of a continuous function by polynomials: given any finite number of values (x_i, y_i), you can always find a polynomial passing through them. However, there are continuous functions which are nowhere differentiable and which have unbounded variation between x_i and x_{i+1} for each i. The "polynomials" are the "nice images" and the "continuous with unbounded variation" are the "noisy ones". Disclaimer: I am neither a lawyer not an expert on NN. Edit: I would like to be corrected if I the similitude is wrong.