Understanding the Role of Individual Units in a Deep Network

https://dissect.csail.mit.edu/

"A computer could draw a scene in two ways:
1. It could compose the scene out of objects it knows.
2. Or it could memorize an image and replay one just like it.
Our GAN Paint demo and our GAN Dissection method provide evidence that the networks have learned some aspects of composition.
[...]
When a deep network is trained on a high-level task such as classifying a place or synthesizing a scene, individual neural units within the network will often emerge that match specific human-interpretable concepts, like "trees", "windows", or "human faces."

What role do such individual units serve within a deep network?

We examine this question in two types of networks that contain interpretable units: networks trained to classify images of scenes (supervised image classifiers), and networks trained to synthesize images of scenes (generative adversarial networks)."
#gan

BSP-Net: Generating Compact Meshes via Binary Space Partitioning

We present BSP-Net, a network that generates compact meshes via binary space partitioning.
Our generated meshes are low-poly and guaranteed to be watertight.
The edges can reproduce sharp details, yet can still approximate curved boundaries.

* github
* paper
* project page
#furniture_reconstruction #3d