# SynFlow: Pruning Neural Networks without data An interesting note: By defining the concept of a **saliency score**, we can then derive two **conservation laws** which allow us to focus only on the high level graph structure of the network, without needed to touch the data at all. This is a general concept that I referrence in [Avoid touching the Data Layer](Avoid%20touching%20the%20Data%20Layer.md). Another interesting note: we may be able to *constructively build networks that observe the properties we are interested in* (see [here](https://youtu.be/8l-TDqpoUQs?t=2646)). We should be able to derive principles of what we want in how the weights are structured and then construct networks according to that. --- Date: 20211115 Links to: [Neural Networks MOC](Neural%20Networks%20MOC.md) Tags: References: * [Youtube paper review](https://www.youtube.com/watch?v=8l-TDqpoUQs&t=2s)