Sophie Zhao·Oct 20One Arrow, Four Roles — A Geometric Picture of Neural Networks and BackpropagationMost textbooks teach neural networks in separate stages. First comes the forward pass: multiply by a weight matrix and add a bias. Then…A response icon1A response icon1
Sophie Zhao·Aug 13Why Attention Needs Linear Layers for Q, K, and V — How Linear Projections Work InternallyIn the transformer architecture, the linear layers transform a normalized copy…A response icon4A response icon4
Sophie Zhao·Jun 21Why Transformer Representations Tend to Live on HyperspheresCosine similarity, angular distance, vector directions, activation steering, and representation arithmetic are everywhere in modern deep…A response icon8A response icon8
Sophie Zhao·Jan 15Why an Inner Product Automatically Gives Length and Angles in Riemannian GeometryOne thing that confuses many people in Riemannian geometry is this:A response icon3A response icon3
Sophie Zhao·Nov 9, 2025Do You Think I’m in Your Palm, or You’re in My Dream?If AI Escapes from the Five-Finger Mountain.A response icon1A response icon1
Sophie Zhao·Oct 1, 2025The Phoenix’s Blueprint: Signs You Have Risen from the Ashes of the Dark Night of the SoulAt some point, we encounter a crisis so deep that it shakes the very core of the soul — a season marked by pain, confusion, and shadow…
Sophie Zhao·Sep 27, 2025Tracking Matrix Shapes Inside a TransformerTransformers have taken over deep learning — from BERT to GPT — but when you open the hood, the matrix and tensor shapes can feel slippery.
Sophie Zhao·Sep 24, 2025Why Deep Learning Loves the Dot Product and Matrix MultiplicationThe dot product is more than a basic linear algebra operation — it is one of the core mathematical tools that powers modern deep learning…A response icon2A response icon2
Sophie Zhao·Jul 21, 2025Three Ways to Solve Bellman Equations: Intuition Behind Policy Iteration, Value Iteration, and…In this article, we’ll intuitively explore three major ways to solve Bellman Equation: Policy Iteration, Value Iteration and Q-Learning.A response icon1A response icon1
Sophie Zhao·Jul 21, 2025What is Bootstrapping in Reinforcement Learning?Ha! ‘Bootstrap’ ? What does it mean?A response icon1A response icon1