in Seminar ~ read.

Seminar 12.1 Review

Deep image prior

arXiv:1711.10925

This paper gives a no learning image reconstruction method using a neural network.

In stead of the traditional regularization term like TV, wavelet and etc. This paper consider "The likelihood of the image can be reconstructed by an encoder-decoder like neural network" as a regularization.

Architechture: an encoder-decoder like neural network like U-net

Prons:

(+) To best of my knowledge, this image reconstruction is the sate-of-the-art method to reconstruction by a single image[beats BM3D, DDTF etc.] It is also surprising that this algorithm also beats lots of the state-of-the-art learning algorithm.

Cons:

(-) The computational cost is very high, so this algorithm may not useful.

Further Directions:

  1. A Better regularization term design, this paper gives us confidence that

  2. How to utilize the encoder-decoder network after the training?

  3. faster algorithms, like using a image generator.

Dynamic Routing Between Capsules

arXiv:1710.09829

confidence: I'm not quite sure I have understood this paper

Motivation: Max pooling loose information of position relationship?

Best of my knowledge, Here are several works that also can utilize this information like some recent paper in video/image detection

  1. Relation Networks for Object Detection: https://arxiv.org/pdf/1711.11575 [use self attention in object detection]
  2. Nonlocal neural network: https://arxiv.org/pdf/1711.07971

I need to read carefully about the part of dynamic routing methods and the motivation!

Experiment

  1. adding a generator/decoder network as regularization

  2. Individual dimensions of a capsule represent in section 5.1

  3. better generalization:Robustness to Affine Transformations

  4. Maybe i need to read carefully about the MultiMNIST results!