Criticisms of Machine Learning / Deep Learning

Criticisms of Deep Learning

Francois: https://blog.keras.io/the-limitations-of-deep-learning.html Gary Marcus: Rebooting AI Judea Pearl: Theoretical Impediments to Machine Learning https://arxiv.org/abs/1801.04016 Hector Zenil: https://www.quora.com/What-are-the-main-criticism-and-limitations-of-deep-learning Zach Liption: Deep Flaws In Deep Learning https://www.kdnuggets.com/2015/01/deep-learning-flaws-universal-machine-learning.html Deep Networks are Easily Fooled: https://arxiv.org/pdf/1412.1897v2.pdf Boston Dynamics uses no learning, rather Control Theory Next to no real world RL applications.

Generator of criticisms: Take any slowly moving subfield of machine learning research, and claim that fundamentally new ideas are necessary to accomplish what the subfield is trying to accomplish (the subfield’s existence means that there are unsolved problems). Subfields: https://docs.google.com/document/d/1G-ppYPhrAm82PqJWidx75EjvMDz8SRLvbC9QXuIq17k/edit?usp=sharing

Stronger generator: Take problems that are currently better solved by methods outside machine learning and demonstrate / claim that learning does not and will not solve them well. Research subfields in broader computer science.

Shortlist:

Causality
Logic / Reasoning (Deduction)
Symbol Discovery / Abstract Knowledge
High Data Requirements (for DL)
Poor transfer
Differentiability / Smoothness requirements

This is an argument for discrete latents.

Source: Original Google Doc