Have you ever wondered how scientists create holograms for atomic experiments? Recently, I stumbled upon a fascinating paper that used a neural network to compute holograms. As a neural network newbie, I was intrigued and decided to dive deeper. In this post, I’ll share my journey of building a neural network for computing holograms, including the challenges I faced and what I learned along the way.
The paper I referenced used residual blocks in the upsampling path, which was confusing at first. I didn’t know how to implement the downsampling and upsampling blocks or how to design the bottleneck and the first and last convolutional layers. After some research, I came up with the following architecture:
– Downsampling block: Conv 4×4 (stride=2, Padding=1)+ReLU+BatchNorm2D
– Residual Block: (full preactivation+identity skip): BatchNorm2D+ReLU+Conv 4×4 (stride=1, padding=2) x2
– Upsampling block: TConv 4×4 (stride=2, Padding=1)+BatchNorm2D+ReLU
However, I’m still unsure if this architecture makes sense. I’d love to hear from experts in the field or anyone who has experience with neural networks.
Computing holograms is an exciting area of research, and I believe that neural networks can play a significant role in advancing this field. If you’re interested in learning more about this topic or have any insights to share, please leave a comment below.