Lecture - Theories of Deep Learning MT25, XII, Vulnerabilities in deep learning models

[[Course - Theories of Deep Learning MT25]]^U

• Given a task, consider all the invariants you want to model and then design the activation functions to remove these. The scattering transform does this for translation

• Principal component analysis can be viewed as an autoencoder
• Autoencoders can be viewed as an extension of PCA by allowing transformations that aren’t just low dimensional projections
• $k$-sparse autoencoders:

• How do the decision region diagrams change when you have an additional “don’t know” class?