4 ms·
Indeed. I had a light-bulb moment about this when thinking about how a 2D image is unrolled into a one-dimensional vector before processing by an artificial neu
by wcoenen 9y ago
Indeed. I had a light-bulb moment about this when thinking about how a 2D image is unrolled into a one-dimensional vector before processing by an artificial neural network. Before training, the NN doesn't even have a concept of certain pixels being related because of proximity. It needs to learn to see the 2D structure in the data from scratch, starting with edge detection.
And this also means that the order in which the pixels are presented to the NN doesn't matter in the first place. You could pick a completely random (but fixed) permutation and the NN would still learn to "see" in that mess.
- llukas 9y agoNN has concept of 2D image built in - conv layer for example: http://cs231n.github.io/convolutional-networks/ http://cs231n.github.io/convolutional-networks/
- wcoenen 9y agoThanks, I don't think CNNs were mentioned in Andrew Ng's machine learning course. That's about the only background I have in machine learning. The NN examples of image recognition in that course did start from scratch without any architectural features designed for images.