Visual Lab — Overview
Every other page on this site explains a concept, and most back that explanation with a generated chart or animation. This section is different: every visualization here is live — you set the parameters, you click the canvas, you watch it respond in real time. Reading about gradient descent tells you what should happen; dragging the starting point and watching Adam actually out-turn SGD around a ravine is a different, stronger kind of understanding.
How to Use This Section
Each page below is one interactive component, with a short "what to try" list. Most are also embedded directly on the content page they belong to (the Gradient Descent Explorer lives on both this page and Optimizers, In Full Depth, right where the static version used to be alone) — this section exists so every interactive visualization also has one permanent, browsable home, independent of which content page you started from.
What's Here So Far
- Gradient Descent Explorer — click to set a starting point on a real loss surface, choose SGD/Momentum/Adam, and watch each one converge (or not) live.
- Neural Network Playground — a real, hand-written neural network training live on a toy dataset you pick, with full control over depth, width, and activation.
- Attention Step-Through — type a sentence, watch a real attention heatmap and per-token breakdown compute live, across multiple heads.
- RAG Pipeline Simulator — a real, clickable retrieval pipeline: chunking, cosine-similarity ranking, and reranking, responding live to whatever you ask.
- Agent Execution Graph — step through a scripted ReAct loop one Thought/Action/Observation at a time, or watch it play automatically.
- LLM Inference Flow Visualizer — the text-to-token pipeline as a clickable graph, with a real live temperature/top-k/top-p sampling playground.
- Decision Boundary Playground — click to place points, pick KNN / decision tree / k-means, and watch a real boundary recompute live.
- Embedding Space Explorer — a real from-scratch PCA projection, cosine-similarity nearest neighbors, and the king − man + woman ≈ queen analogy, live.
- Linear Regression Studio — drag the line yourself, or hand it to real gradient descent and watch it converge (or diverge) on a real dataset, loss surface included.
- Logistic Regression Studio — a real sigmoid fit, plus a live side-by-side of cross-entropy vs. the non-convex MSE-on-sigmoid trap, both computed from scratch.
- Ridge Regression Studio — a real closed-form ridge solve on a deliberately multicollinear dataset; watch unstable OLS coefficients settle down as λ increases.
- Lasso Regression Studio — real coordinate-descent soft-thresholding on the same dataset as Ridge; watch coefficients hit exactly zero instead of just shrinking.
- Algorithm Selector — answer a few questions about your problem and get a real algorithm recommendation, linked to this site's own guide for it.
- Two Pointers & Sliding Window, Step by Step — edit an array or string and step through the real two-pointer or sliding-window algorithm one move at a time, the exact logic behind two of the Practice Problems.
- Graph & Tree Traversal, Step by Step — edit a graph and step through real BFS vs. DFS, or pick any two nodes on a fixed tree and watch the real recursive Lowest-Common-Ancestor search unfold.
See the roadmap for what's shipped and what might come next.