Part 1 Contents

The Machine Learning Pipeline

Every ML system follows this basic flow: collect data, train a model, then use it to make predictions.

Model

A program that has learned patterns from data. Once trained, you can feed it new inputs and it will generate predictions based on patterns it discovered. When people say "GPT-4" or "Claude," they are referring to specific models.

Analogy: An experienced chef's accumulated cooking knowledge. Give them new ingredients, and they instinctively know what to make.

Training

The learning phase where a model sees millions or billions of examples and gradually adjusts its internal numbers to find patterns. Can take days or weeks on powerful hardware and is typically the most expensive part of building AI.

Analogy: Practicing 10,000 basketball free throws. Each shot gives feedback. Over time, your brain automatically adjusts your form.

Inference

Using a trained model to make predictions. When you ask ChatGPT a question, it performs inference. Requires less compute than training, but when millions use a service, inference costs add up quickly.

Analogy: Training is studying for an exam over weeks. Inference is sitting down and actually taking the exam.

Parameters: The Model's Adjustable Knobs

Large models have billions or trillions of these adjustable values. More parameters generally means more capacity to learn nuance.

Parameters

The adjustable numbers inside a model. During training, these get tuned to capture patterns in the data. A model with 7 billion parameters has 7 billion adjustable numbers. More parameters means more capacity to learn complex patterns.

Analogy: Sliders on an audio mixing board. A basic board has 8 sliders. A professional board has 128 for precise control.

Weights

The learned values stored in parameters. When someone says "download the model weights," they mean downloading everything the model learned. Without weights, you have an empty architecture with no knowledge.

Analogy: If parameters are sliders, weights are where each slider is positioned after careful tuning.

The Training Loop

Loss

A number measuring how wrong the model's predictions are. The entire goal of training is to minimize this. Lower loss means better predictions. Training graphs show loss decreasing over time.

Analogy: Your golf score. Lower is better. Each training step tries to shave strokes off your game.

Gradient Descent: Finding the Lowest Point

Gradient

Tells you which direction to adjust each parameter to reduce loss. Mathematically, it is the slope of the loss function. The gradient points uphill, so you go the opposite direction.

Analogy: Standing blindfolded on a foggy mountain. You feel which way slopes down—that feeling is the gradient.

Gradient Descent

The algorithm that repeatedly follows gradients to find the lowest loss. Calculate gradient, take a step, repeat. SGD and Adam are popular optimized versions.

Analogy: Walking downhill step by step until you reach the valley floor.

Learning Rate

Controls how big each step is during gradient descent. Too large and you overshoot the optimal point. Too small and training takes forever. Finding the right balance is crucial.

Analogy: Your stride length walking downhill. Giant leaps might overshoot; tiny steps take all day.

Forward Pass and Backward Pass

Forward: Data flows through the network to produce a prediction. Backward (backpropagation): Error flows backward to calculate how much each parameter contributed to the mistake.

Analogy: Forward is a factory assembly line producing a product. Backward is quality inspectors tracing a defect back through each machine.

Epoch and Batch

Epoch: One complete pass through the entire dataset. Training typically runs for many epochs. Batch: A group of examples processed together before updating weights (typically 32 or 64).

Analogy: Epoch is reading a textbook cover to cover. Batch is grading 32 papers at once before giving feedback.

Overfitting

When a model memorizes training data instead of learning general patterns. Performs excellently on training data but poorly on new data. One of the most common problems in ML.

Analogy: A student who memorizes test answers without understanding. Aces practice tests but bombs the real exam.

A Simple Neural Network

Each circle is a neuron. Each column is a layer. The arrows represent weights connecting neurons.

Neuron

The basic unit in a neural network. Takes inputs, multiplies each by a weight, adds them with a bias, and passes the result through an activation function. Named after biological neurons, though the analogy is loose.

Analogy: A voter weighing different factors to make a decision. Many neurons together produce intelligent outputs.

Layer

A group of neurons that process data together. Networks stack layers: input receives data, hidden layers process it through multiple stages, output produces the result. Modern LLMs have over 100 layers.

Analogy: Floors in a building. Data enters ground floor, gets processed on each floor, final decision at the top.

MLP (Multi-Layer Perceptron)

The simplest neural network: fully connected layers stacked together. "Fully connected" means every neuron in one layer connects to every neuron in the next. MLPs appear as components inside transformers.

Analogy: A game of telephone where everyone on one level talks directly to everyone on the next level.

Common Activation Functions

Activation Function

Applied after each neuron to introduce non-linearity. Without them, networks could only learn straight-line relationships. Common choices include ReLU, GELU, sigmoid, and tanh.

Analogy: A threshold for firing. Like a biological neuron that only fires if the signal is strong enough.

ReLU (Rectified Linear Unit)

The most popular activation function. Rule: if negative, output zero; if positive, pass through unchanged. Simple but remarkably effective. GELU is a smoother variant used in transformers.

Analogy: A nightclub bouncer who only lets positive vibes in. Negative energy? You are out.

Softmax

Converts a list of numbers into probabilities that sum to 1. Used at output layers when picking one option from many. In language models, softmax converts scores for each possible next word into probabilities.

Analogy: Converting race finish times into "probability of winning." The winner gets emphasized.