This page contains errata for the paper “Xiang Zhang, Junbo Zhao, Yann LeCun. Character-level Convolutional Networks for Text Classification. Advances in Neural Information Processing Systems 28 (NIPS 2015)“. The paper on arXiv server will be updated accordingly, but the paper in NIPS proceedings may stay as is. Some of the errata may also apply to our earlier technical report “Xiang Zhang, Yann LeCun. Text Understanding from Scratch. arXiv 1502.01710.”
The upper index for the convolution and max-pooling module should be \( \lfloor (l-k)/d \rfloor + 1 \) instead of \( \lfloor (l-k+1)/d \rfloor \).
Today I saw a post from Mark Zuckerberg on his project of automating a home. That great project is so awesome that it could probably change the future of human living. Along with the post there are arguments that common sense may be obtained through unsupervised learning. But personally I do not buy these arguments, at least not in the current form of unsupervised learning.
The intuition that a model can obtain common sense through unsupervised process is not straightforward to me.
Here is one simple question: how to make machines learn to sum up two numbers? Of course, this problem largely depends on how the numbers are represented. If they are represented in some finite-precision float-point format, a simple regression where both weights are one would solve the problem. But that’s not what I mean here. What I mean is, given the symbolic representation of numbers (i.e., each number is a sequence of digits), how could a machine learn to sum them up?
Update June 8th 2015: The dataset duplication issues are fixed in the latest revision of our technical report. Some of our large-scale datasets became smaller than before, but the general conclusion in the technical report still holds. The information below is retained for your reference, although they are no longer valid. We are working on extending comparisons with stronger baseline models and releasing the datasets as soon as possible.