A mental framework for understanding, explaining, and reasoning about any ML/DL algorithm by decomposing it into two core threads: forward (data→prediction) and backward (learning/correction). Covers ALL ML families: deep learning (Dense, CNN, RNN, Transformer, Attention), classical ML (KNN, SVM, Decision Tree, Random Forest, Naive Bayes, Linear/Logistic Regression), unsupervised (K-Means, PCA, DBSCAN), and ensemble methods (Bagging, Boosting, Stacking). Use when asked to explain any ML algorithm, layer, or architecture — or when answering ML/DL interview questions, teaching, debugging model behavior, or building intuition. Triggers: ML/DL explanation, algorithm analysis, interview prep, "how does X work", model walkthrough, gradient flow, teaching/tutoring.