Classifying Images: Massive Parallelism And Surface Features

Analysis of image classifiers demonstrates that it is possible to understand backprop networks at the task-relevant run-time algorithmic level. In these systems, at least, networks gain their power from deploying massive parallelism to check for the presence of a vast number of simple, shallow patterns.

https://betterwithout.ai/images-surface-features

This episode has a lot of links:

David Chapman's earliest public mention, in February 2016, of image classifiers probably using color and texture in ways that "cheat": twitter.com/Meaningness/status/698688687341572096

Jordana Cepelewicz's "Where we see shapes, AI sees textures," Quanta Magazine, July 1, 2019: https://www.quantamagazine.org/where-we-see-shapes-ai-sees-textures-20190701/

"Suddenly, a leopard print sofa appears", May 2015: https://web.archive.org/web/20150622084852/http://rocknrollnerd.github.io/ml/2015/05/27/leopard-sofa.html

"Understanding How Image Quality Affects Deep Neural Networks" April 2016: https://arxiv.org/abs/1604.04004
 
Goodfellow et al., "Explaining and Harnessing Adversarial Examples," December 2014: https://arxiv.org/abs/1412.6572

"Universal adversarial perturbations," October 2016: https://arxiv.org/pdf/1610.08401v1.pdf

"Exploring the Landscape of Spatial Robustness," December 2017: https://arxiv.org/abs/1712.02779

"Overinterpretation reveals image classification model pathologies," NeurIPS 2021: https://proceedings.neurips.cc/paper/2021/file/8217bb4e7fa0541e0f5e04fea764ab91-Paper.pdf

"Approximating CNNs with Bag-of-Local-Features Models Works Surprisingly Well on ImageNet," ICLR 2019: https://openreview.net/forum?id=SkfMWhAqYQ

Baker et al.'s "Deep convolutional networks do not classify based on global object shape," PLOS Computational Biology, 2018: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1006613

François Chollet's Twitter threads about AI producing images of horses with extra legs: twitter.com/fchollet/status/1573836241875120128 and twitter.com/fchollet/status/1573843774803161090

"Zoom In: An Introduction to Circuits," 2020: https://distill.pub/2020/circuits/zoom-in/

Geirhos et al., "ImageNet-Trained CNNs Are Biased Towards Texture; Increasing Shape Bias Improves Accuracy and Robustness," ICLR 2019: https://openreview.net/forum?id=Bygh9j09KX

Dehghani et al., "Scaling Vision Transformers to 22 Billion Parameters," 2023: https://arxiv.org/abs/2302.05442

Hasson et al., "Direct Fit to Nature: An Evolutionary Perspective on Biological and Artificial Neural Networks," February 2020: https://www.gwern.net/docs/ai/scaling/2020-hasson.pdf