Part of AGI experiments
Domain Quantization
Information-density normalization - adjust histogram bins so precision follows where the data is.
Each input is represented as a histogram, and the model adjusts the bin boundaries so that they pack tightly where data points are dense and spread out where they are sparse - precision following information density. The effect is a learned transform that maps an input drawn from the expected distribution onto a roughly piecewise-linear output, which can make the job of downstream layers considerably easier. It is a normalization idea grounded in how much information actually lives in each region of the input space.