Neural Networks Basics

Introduction

A neural network is layers of simple math units ("neurons") that learn to transform inputs into outputs by adjusting weights during training. They power every modern AI — including the LLMs inside agents.

Beginner analogy: imagine 1,000 friends voting on whether a picture is a cat. Each friend has a slightly different opinion, weighted by how often they're right. Training = adjusting whose vote counts more, until the group is accurate.

Understanding the topic

Core concepts:

• Neuron: weighted sum of inputs → activation function → output.
• Layers: input → hidden → output; deep = many hidden layers.
• Training: gradient descent on a loss function.
• Backprop: how errors flow backwards to update weights.
• Transformer = neural net specialised for sequences (text).

Syntax reference

Visual workflow / architecture:

input ─►[neuron][neuron][neuron]─► hidden
           │       │       │
           ▼       ▼       ▼
        [neuron][neuron][neuron]
                  │
                  ▼
              output / prediction
                  │
            loss = error
                  │
           backprop updates weights

Real-world use

Neural nets power image classification (ResNet), speech (Whisper), translation (Google Translate), and text generation (GPT, Claude, Llama).

Best practices

• Understand the maths at a high level; you don't need to derive backprop daily.
• Use pretrained models — training from scratch is rarely worth it.

Common mistakes

• Thinking you must train models to build AI products — 99% of agent work uses APIs.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is a neuron in machine-learning terms?
• Q2. Why is backprop important?
• Q3. What's the difference between training and inference?
• Q4. Why are transformers special for text?