differentiability
The mathematical property of neural networks that lets their many parameters be smoothly and continuously tuned via gradients; Tegmark argues this, not inscrutability, is the source of their power.
As defined in 75 | Max Tegmark on Reality, Simulation, and the Multiverse
Where it comes up · 1
Most important claims · 1 of 1Search all claims for “differentiability” →
- It is a common misconception that the power of neural networks comes from their inscrutability; in fact their power comes from their differentiability—being a black box with many knobs that can be continuously tuned until they perform well—so it should be possible to combine deep learning's ability to discover solutions with a subsequent step that simplifies those solutions into human-understandable, intelligible forms, exactly as a human can both learn to catch a ball and then explain its parabolic path with an equation.75 | Max Tegmark on Reality, Simulation, and the Multiverse