Reading Assignment: Advanced CNN Architectures

CNN architecture overview

(if you have not read it yet)

Start with the high-level overview provided in this blog post by Adit Deshpande. Anything beyond the section on region-based CNNs is optional. (GANs are covered in the “Learning Generative Models” course.)

Next, read Densely connected convolutional networks (2016), Huang et al. There is a lot to learn from this paper as the authors do a very good job pointing out similarities and differences with many related approaches.

Detailed Reading

Here are 6 papers that use (and extend) CNNs in various settings.
They are all worth reading, but you don’t need to read all of them.
Each of you needs to choose one paper to explain it to the others next week.
Please use this Doodle poll for choosing a paper (max 3 people for one paper).
In case somebody cannot make it to the class next week: Don’t do the poll, so we have our groups as complete as possible.
I expect 18 slots to be enough, but if all slots happen to be full, please contact me!

Here are some questions to guide you during reading:

• What is the learning problem? (datasets, representation of inputs and outputs, cost function and evaluation measures) Which specific challenges are addressed?
• What does the network design look like? (Try to understand the details as if you were to implement the described architecture!)
• What are the main innovations described in the paper? Where does it go beyond the techniques covered in the deep learning book?
• What are the main experimental findings?
• Does the paper have any weak spots?
• Is there any information on performance (throughput) and hardware requirements?

1. A convolutional neural network for modeling sentences (2014), N. Kalchbrenner et al.

2. End-to-end Learning for Music Audio Tagging at Scale (2018), J. Pons et al.

3. Image style transfer using convolutional neural networks (2016), L.A. Gatys et al.

4. Image Super-Resolution Using Deep Convolutional Networks (2016), C. Dong et al.

5. Singing voice separation with deep U-Net convolutional networks (2017), A. Jansson et al.

6. You only look once: Unified, real-time object detection (2016), J. Redmon et al.

Optional Further Reading

If you would like to dig deeper, here are some more resources:

• on ResNets: Deep residual learning for image recognition (2016), K. He et al.
• on Inception architectures: Rethinking the inception architecture for computer vision (2016), C. Szegedy et al.
• on 1x1 convolution: One by One [ 1 x 1 ] Convolution - counter-intuitively useful (2016), A. Prakash
• on convolution arithmetic A guide to convolution arithmetic for deep learning (2016), V. Dumoulin & F. Visin