Lesson 2: How Neural Networks Learn (1 hour)
Learning Objectives
- Understand the learning process in neural networks
- Understand the concept of weights and how they're adjusted
- Understand the basic idea of backpropagation (conceptually)
- See how training improves network performance
Materials Needed
- Internet connection
- Neural network training visualization
- Examples of learning process
- Student notebooks
- Whiteboard for diagrams
Time Breakdown
- Review neural network structure (5 min)
- Introduction to learning (15 min)
- Weights and adjustments (15 min)
- Training process visualization (20 min)
- Wrap-up (5 min)
Activities
1. Review Neural Network Structure (5 min)
- What are the three main parts? (Input, hidden, output layers)
- What connects neurons? (Connections with weights)
- Bridge: "How do networks learn the right weights?"
2. Introduction to Learning (15 min)
The Learning Problem:
- Neural network starts with random weights
- Doesn't know anything yet
- Needs to learn correct weights to make good predictions
The Learning Process:
- Feed forward: Input data flows through network → prediction
- Compare: Compare prediction to correct answer → error
- Backpropagate: Send error back through network
- Adjust weights: Change weights to reduce error
- Repeat: Do this many times with many examples
Simple Analogy: Learning to Throw a Ball
- First attempts: Way off target (random weights)
- Each throw: See how close you got (compare prediction)
- Adjust: Change how you throw (adjust weights)
- Many attempts: Get better (training)
- Eventually: Hit target consistently (learned)
Key Concept:
- Network makes a guess
- Sees how wrong it was
- Adjusts to be less wrong
- Repeats until good enough
3. Weights and Adjustments (15 min)
What are Weights?
- Numbers that determine connection strength
- Like: How much does this input matter?
- Example:
- Input: "Has fur" (weight: 0.8 - very important)
- Input: "Color" (weight: 0.3 - less important)
- Input: "Size" (weight: 0.5 - somewhat important)
Initial Weights:
- Start random (network knows nothing)
- Like: Guessing randomly
Learning = Adjusting Weights:
- If network's prediction is wrong:
- Increase weights that would have helped
- Decrease weights that caused mistakes
- Over time, weights converge to good values
Learning Rate:
- How much to adjust weights
- Too small: Learns very slowly
- Too large: Might overshoot, never converges
- Just right: Learns efficiently
Visual Example:
- Show simple network with weights
- Show one training example
- Show how weights adjust
- Show how prediction improves
4. Training Process Visualization (20 min)
Activity 1: Watch Training in Action (10 min)
- Use online neural network training visualization
- Show network learning to classify images or recognize patterns
- Observe:
- Starting accuracy (random, ~50%)
- How accuracy improves over time
- How weights change
- How network gets better with each example
Activity 2: Human Learning Simulation (10 min)
- Students act as neural network
- Setup: 3 students as input neurons, 2 as hidden, 1 as output
- Task: Learn to recognize "happy" vs. "sad" faces
- Process:
- Show input (face description)
- Each neuron makes decision based on weights
- Output neuron makes prediction
- Teacher gives correct answer
- Students adjust their "weights" (how they decide)
- Repeat with new example
- Observe: Gets better over time!
Key Observations:
- Starts with random guesses
- Gets better with each example
- Eventually makes good predictions
- More examples = better learning
5. Wrap-Up (5 min)
- Learning = adjusting weights to reduce errors
- Process: Predict → Compare → Adjust → Repeat
- More training = better performance
- Preview: Next lesson - Deep learning and why depth matters
Differentiation Strategies
- Younger students: Focus on analogies, simpler explanations, hands-on simulation
- Older students: Introduce gradient descent concept, explore learning rates, research backpropagation
- Struggling learners: Use physical simulation, simpler examples, more repetition
- Advanced learners: Research optimization algorithms, explore different loss functions, analyze training dynamics
Assessment
- Understanding of learning process
- Participation in activities
- Quality of observations
- Reflection journal entry