DDPM learns generation by reversing Gaussian noise

A theory-first walkthrough of Denoising Diffusion Probabilistic Models. The post builds from the fixed forward corruption process to the learned reverse process, derives the noise-prediction objective, explains its connection to score matching and Langevin-like sampling, and shows why a timestep-conditioned U-Net can generate images from noise.

29 min read   ·   August 02, 2026

2026   ·   ddpm   diffusion   generative-models   u-net   score-matching     ·   paper-notes  

Attention is an information-routing operator

A conceptual walkthrough of attention as learned information routing. The post builds from tokens and Q/K/V projections to self-attention, multi-head attention, masking, positional encoding, and how the same ideas appear in robotics, VLMs, and VLA-style policies.

15 min read   ·   June 29, 2026

2026   ·   attention   network-architecture   transformers   self-attention     ·   paper-notes  

Understanding U-Net as a “what → where” machine

A visual walkthrough of U-Net for image segmentation, explaining how CNNs learn “what” is in an image but lose precise “where” information. The post breaks down the U shaped architecture in down-sample blocks, up-sample blocks, skip connections, transposed convolutions, and supervised training with segmentation masks.

13 min read   ·   June 26, 2026

2026   ·   u-net   network-architecture   segmentation   deep-learning     ·   paper-notes