AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
The Problem of Information Plasticity in Model Compression
The architecture, what i'm taking to be the structure of the models, matters a lot for the reasons that we've talked about. We have the concept that we call information plasticity as analogous to neuronal plasticity. So it appears that the effective conductivity, which is not the physical connectivity thats determined at the outset, but the effective connectivity, as measured by the amount of information thaf that connection carries. And if you knew that at the beginning, you could train with a smaller network, the promise that to get there, you need to go through this very high bump an information in the network.