https://youtu.be/Mljfam5zNJo?t=66 - Theme: Architecture
Densely Connected Convolutional Networks
https://arxiv.org/abs/1608.06993
Best paper award (Facebook)
tl;dr; Connect the conv layer not only with the next layer but also with some layers after it.


https://youtu.be/Mljfam5zNJo?t=484
Annotating Object Instances with a Polygon-RNN
https://arxiv.org/abs/1704.05548
tl;dr; conv LSTM that predicts a sequence of points (possibly starting with a user provided point)

https://youtu.be/Mljfam5zNJo?t=1036 - Theme: Vision tasks
PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
https://arxiv.org/abs/1612.00593
tl;dr; [classify/segment point cloud, need to be able to rearange or rotate points]
       use operations that either ignore permutations in input or permute out in a same way

https://youtu.be/Mljfam5zNJo?t=1494
FCSS: Fully Convolutional Self-Similarity for Dense Semantic Correspondence
https://arxiv.org/abs/1702.00926
tl;dr; [deep learned sift replacement] no arch review

https://youtu.be/Mljfam5zNJo?t=1762
Full Resolution Image Compression with Recurrent Neural Networks
https://arxiv.org/abs/1608.05148
no arch review

https://youtu.be/Mljfam5zNJo?t=1852
Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
https://arxiv.org/abs/1609.04802
(Twitter)
no arch review

https://youtu.be/Mljfam5zNJo?t=1997
One-to-Many Network for Visually Pleasing Compression Artifacts Reduction
https://arxiv.org/abs/1611.04994
tl;dr; [recover image after JPEG compression] add an aditional loss that recovered image should match input when compressed via JPEG

https://youtu.be/Mljfam5zNJo?t=2307
Network Dissection: Quantifying Interpretability of Deep Visual Representations
https://arxiv.org/abs/1704.05796
tl;dr; [interpret how the network sees]
   a) build a heatmap of pixels that activate some neuron in the middle of the network
   b) compare this heatmap with the ground truth for the segmentation task.
   c) if the heatmap and the GT match for a specific class then mark this neuron as the one that activates for this class
   Apparently there are quite a lot of neurons like this

https://youtu.be/Mljfam5zNJo?t=2584
Training Sparse Neural Networks
https://arxiv.org/abs/1611.06694
Local Binary Convolutional Neural Networks
https://arxiv.org/abs/1608.06049

https://youtu.be/Mljfam5zNJo?t=2655
Universal Adversarial Perturbations
https://arxiv.org/abs/1610.08401
tl;dr; [find a universal image (~noise) that you can add to any image and it will lead to misclassification]