What makes ImageNet good for transfer learning?

Minyoung Huh, Pulkit Agrawal, Alexei A. Efros. arXiv 2016

Presented by: Ismail

Lets recap..

Week 1

So far..(What is ImageNet?)

So far.. (What is AlexNet?)

So far..(Performance of AlexNet!)

So far..(CNN activation as features?)

Slide credit: Huan Zhang, UCDavis

So far..(CNN activation as features?)

Slide credit: Huan Zhang, UCDavis

So far..(Can we do transfer learning?)

Slide credit: Jason Yosinski

Week 2

So far..(CNN features for object detection?)

Slide Credit: Ross Girshick

So far..(Pre-Training?)

Slide credit: Patrick Chen, UC Davis

Are these performance increase restricted to ImageNet?

1. How does the amount of pre-training data

affect transfer performance?

1. How does the amount of pre-training dataaffect transfer performance?

2. How does the taxonomy of the pre-training task affect transfer performance?

Bottom-up: 918, 753, 486, 79 and 9 classesTop-down: 127, 10 and 2 classes

2.1 -- Effect of number of pre-training classes on transfer performance?

Top-down: transfer performance

2.2 -- Does training with coarse classes induce features relevant for fine-grained recognition?

Induction accuracy, top-1 and top-5 NN in FC7

2.3 -- Does training with coarse classes induce features relevant for fine-grained recognition?

Induction accuracy, top-1 and top-5 NN in FC7

2.4 -- Does training with fine-grained classes induce features relevant for coarse recognition?

2.5 -- More Classes or More Examples Per Class?

2.6 -- How important is to pre-train on classes that are also present in a target task?

3. Does data augmentation from non-target classes always improve performance?

Splitting ImageNet..

Does adding arbitrary classes to pre-training data always improve transfer performance?