I wanted to look into filters as "pattern detectors" and found this article which talks about how random filters can act similar to edge detectors. Really neat idea and really emphasizes the idea in the note on this slide.
I took a look at the article you linked and I found it fascinating as well! The idea of convolutions acting as edge detectors reminds me of the convolutional neural networks that were taught in CS 189, where convolutional layers act as filters for higher level features in images (eg. edges).
stephen422
Then can we think of convolutional neural networks as some kind of deeply chained set of pattern detectors? Might be related to why CNNs work so well for images.
JefferyYC
I agree with Stephen's observation. Essentially a CNN is just layers of convolution weights chained with pooling and linear weights. After learning this lecture I realized this edge detection technique has existed long before the application of CNN. However, CNN did make edge detection much more powerful through finding the appropriate convolution parameters through gradient descent (instead of us hardcoding them in this example).
I wanted to look into filters as "pattern detectors" and found this article which talks about how random filters can act similar to edge detectors. Really neat idea and really emphasizes the idea in the note on this slide.
https://cs.stackexchange.com/questions/51866/computer-vision-why-do-random-filters-perform-similar-to-edge-detectors
I took a look at the article you linked and I found it fascinating as well! The idea of convolutions acting as edge detectors reminds me of the convolutional neural networks that were taught in CS 189, where convolutional layers act as filters for higher level features in images (eg. edges).
Then can we think of convolutional neural networks as some kind of deeply chained set of pattern detectors? Might be related to why CNNs work so well for images.
I agree with Stephen's observation. Essentially a CNN is just layers of convolution weights chained with pooling and linear weights. After learning this lecture I realized this edge detection technique has existed long before the application of CNN. However, CNN did make edge detection much more powerful through finding the appropriate convolution parameters through gradient descent (instead of us hardcoding them in this example).