How Machines Learn

By: Saanvi Chugh

How many of you have used ChatGPT, Gemini, or some other LLM before?Today, I will be explaining how machines are able to make accurate predictions.

Too smokey

Too hot

Too little

Neural network = a machine learning model that learns how to turn raw inputs into useful patterns and then uses those patterns to make predictions

Its ancestor is the linear model y = mx + b

Early layers figure out which parts of the input matter most

Middle layers mix those pieces together to recognize bigger ideas (context/tone)

The final layer uses all of that to calculate the probability of possible predictions

Weights = dials that control how strongly each input feature influences the decision

Biases = built-in values that let the neuron “shift” when it decides to activate, even if inputs are weak

By adjusting these values, the network gradually learns to make accurate predictions.

1. Input layer holds input features (X1, X2, X3, ...)

2. Hidden leayers consist of artificial neurons that transform inputs using weights and biases

3. Output layer produces the final prediction

Inputs: "prize" "money" "dear" "hello" "win"

"prize" is given more weight than "hello," meaning it contributes more to the overall prediction

Eventually, the neural network will accurately predict whether an email is spam or not

Back propagation = an algorithm used to learn the correct weights and biases

Compares prediction Y to true Y and uses a loss function to measure the error
At each neuron, the algorithm uses calculus to determine how much each weight and bias contributed to the error

Computer Vision + Medical Imaging

Chatbots/LLMs

Forcasting

Speech Recognition