Post by Puru Kathuria

Everyone says neural networks are complicated. Start from where they actually started. A line. y = mx + b. One input. One output. A relationship that never curves. The problem: a line can't decide. Give it a tumor scan. It returns 0.73. Is that cancer? The line doesn't know. It gives you a number, not an answer. So someone bent it. The sigmoid takes that 0.73 and squashes into a probability. Now 0.73 means "probably yes." 0.12 means "probably no." One line. One decision. One boundary. But real problems don't have one boundary or a linear boundary. A face. A word. A transaction flagged as fraud. None of these are a single yes-or-no. So the decisions got stacked. Layer one finds simple patterns. Layer two finds patterns in those patterns. Deeper layers keep composing them, carving decision regions that no single line could ever create. A neural network isn't a leap from a line. It's a line that learned to bend - a thousand times over.