Part of AGI experiments
Quantized Distribution Learning
Learning latent-variable probability distributions directly, as histograms over quantized bins.
Can a model learn latent-variable probability distributions directly, instead of assuming a parametric form? The Quantized Distribution Auto-Encoder represents each value as a histogram over quantized bins and learns the distribution end to end. Representing values this way sidesteps the precision-weighting machinery that explicit Gaussian latents require - the spread of the histogram already encodes uncertainty - and it became a recurring building block across the AGI experiments.