Metacademy - Differential geometry for machine learning

Riemannian geometry, which allows spaces to be arbitrarily curved and studied from the inside, rather than requiring them to be embedded in some higher-dimensional space.

If you could figure out what patterns of neural activity in a 100-dimension population are fundamental to that population – and which are just recycled combinations of those fundamental patterns – you could explain that neural population with fewer than 100 dimensions.

deep learning is about static translations,