you can reduce dimensionality by using stride>1 during convolution, which I think is equivalent to stride=1 + pooling. But generally dimensionality reduction is a job of a pooling layer (which does downsampling)
#How does a convolutional layer reduce dimensions of a tensor
earnest wave
so yeah if you convolve with multiple kernels, and stride=1, you'll increase your feature dimension. BUT you will do it by using comparably less weights than when using a fully-connected layer (because kernels are small, and you reuse their weights at many locations)
tons of info here https://www.deeplearningbook.org/contents/convnets.html or search for "cnn stride"
earnest wave
haven't used CNNs so much, maybe someone with way more experience can chime in. There is some animations here 🙂 https://github.com/vdumoulin/conv_arithmetic/blob/master/README.md Can you share a link to the source you are referring to?
the meme on that second link is hilarious 😄
so that second image you shared (image.png), that all checks out. Which bit exactly isn't computing for you?
yeah the N really caught me out as well - it does capture all the N, and then the output dim is the number of kernels. All good, happy I helped you help yourself 😄
ashen bramble
You probably already understand this but the convolution as a mathematical function is fascinating with parallels across all of math and more. I love grant sanderson and his videos and this one is a nice example to explain how the dimension reductionality works and more. I’ve worked with CNNs for 4 years and I had never truly understood how convolutions actually worked until this video. https://youtu.be/KuXjwB4LzSA
