Course 4 - Week 3 - Transpose Convolution Question

Hi guys,

After learning about Transpose Convolution, I’ve been reading some ML papers and there I found the following terms as well.

  • Deconvolutional layers
  • Upsampling layers

These layers are used in the ‘second’ part of the architecture, similar to transpose convolution. These layers are also used for keypoint detection architectures and in my view serve the same purpose as the transpose convolution. But I was wondering whether these terms mean different things? Moreover, having learned about the U-Net architecture, I was wondering what the difference between the U-Net architecture is and the hourglass architecture as they seem alike to me.


Hi Lars,

My two cents is that ‘hourglass’ is a generic term that is applicable to many different types of models using an encoder-decoder paradigm, including U-net. You can have a look here or here.

The term ‘deconvolutional layer’ seems unfortunate and would appear to refer to a transposed convolution layer. You can have a look here (particularly the top answer). Upsampling layers differ from transpose convolutions as they do not transpose. See this post.