Somil Singh·Oct 16Deep Learning #14: A correct sampler can still give the wrong learning signalDerive the energy-model likelihood gradient, then measure what a short local chain leaves out on 61 states.
Somil Singh·Oct 15Deep Learning #13: Lower energy still doesn’t tell you the probabilityBuild an energy-based model you can normalize exactly, check its learning gradient and see where a good optimizer still leaves a bad fit.
Somil Singh·Oct 14Deep Learning #12: A flow can reverse perfectly and still learn the wrong thingTwo reversible layers, one likelihood and a gradient check. Part 12 builds the chain on CPU, with full code and measured output.
Somil Singh·Oct 12Normalizing flows: bend a distribution without losing the way backA curved cloud, a reversible map and a density check. Part 11 fits a tiny flow on CPU, with full code and actual output.
Somil Singh·Oct 11Why DDIM can change the sampling path without changing the noise taskHow the DDIM update shares its noise budget, with a derivation and three runnable checks.A response icon2A response icon2
Somil Singh·Oct 10How DDIM takes fewer steps through the same noise prediction problemWhat happens to the samples when 200 reverse evaluations become 20, 5 or just 1?A response icon1A response icon1
Somil Singh·Oct 9Why predicting noise gives a diffusion model a reverse stepThe model never sees the clean point during generation. So where does its next step come from?A response icon1A response icon1
Somil Singh·Oct 8How a diffusion model learns to turn noise into four clustersIf training starts with real data, how can generation start with noise?A response icon2A response icon2
Somil Singh·Oct 7Why a GAN can reject a fake and still teach almost nothingTwo losses want the same outcome. Why does one give a rejected fake almost no gradient?A response icon2A response icon2
Somil Singh·Oct 6How a GAN learns to turn noise into samplesA generator has no paired target, so where does its learning signal come from?A response icon2A response icon2