- From TD(lambda) to GAE: The Forward-Backward View
Explains TD(lambda)'s forward-backward equivalence, how it leads to GAE for advantage estimation in PPO, and how eligibility traces work with O(1) memory.
13 min - BYOL: Bootstrap Your Own Latent
BYOL proves contrastive learning works without negative samples. A student predicts a momentum teacher's output, and the architecture itself prevents collapse.
21 min - MoCo: Momentum Contrast for Visual Representation Learning
MoCo frames contrastive learning as a dictionary look-up using a queue and momentum encoder, enabling scalable unsupervised representation learning.
21 min - SimCLR: Simple Contrastive Learning for Vision
How a minimalist framework — data augmentation, a nonlinear projection head, and a big batch — made self-supervised ResNet match supervised ImageNet.
21 min - On Variational Bounds of Mutual Information
A unified framework connecting MINE, InfoNCE, NWJ, and TUBA — all variational MI estimators with their derivation relationships and bias-variance tradeoffs.
20 min - Contrastive Predictive Coding
How CPC reframes unsupervised learning as predicting the future in a compact representation space, using InfoNCE to separate real futures from random ones.
16 min - Noise Contrastive Estimation
How turning 100,000-way classification into binary classification makes language model training tractable and sparked modern contrastive learning.
17 min - Finally, a Start
Two days of tinkering, an old-school visitor map, and why I still believe in independent blogs.
2 min
Writing
Notes, paper thoughts, and late-night debugging stories.
Page 1 — Showing 8 of 9 posts