[One-page summary] Signal Processing for Implicit Neural Representations (NeurIPS 2022) by Xu et al.

● Summary: Continuous convolution filter can be uniformly approximated by a linear combination of high-order differential operators

 

● Approach highlight

• Periodic activation function: In implicit neural representation, using the periodic activation function can be useful for models that need to produce the same output for different inputs. Ex) input x, y and output R, G, B values
• Generate new INR(implicit Neural Representations) that combine the derivatives of INR. It can be later decoded
  into discretized forms such as image pixels.

●  Main Results

 

● Discussion

• how many times do we differentiate
• Cost of differential