ML Simulations for Learning (www.shirbhate.com)

A new collection of interactive ML simulations and visualizations has been released to help learners and practitioners build intuition about core algorithms. The suite includes step-by-step visualizations of the Adam optimizer on a simple 2D function and on higher-dimensional functions, an exploration of optimization over a randomly generated challenging terrain (to expose local minima and saddle behavior), an interactive tool demonstrating Byte-Pair Encoding (BPE) tokenization, and a pair of feed‑forward neural network simulators — a basic from-scratch version and a more feature-rich simulator with additional options and a compact view. These tools are significant because they make abstract behaviors tangible: you can watch Adam’s adaptive moment estimates and parameter trajectories, see how landscape geometry causes stagnation or fast convergence, and experiment with different hyperparameters to observe their effects in real time. The BPE visualizer clarifies subword merge mechanics that influence tokenization granularity and model input representation. The NN simulations expose forward passes, activation patterns and weight updates, useful for debugging architectures, teaching backpropagation, or prototyping ideas before coding them in a framework. Overall, this hands-on collection is a practical resource for students, educators and researchers seeking better intuition about optimization dynamics, tokenization effects, and basic neural network mechanics.