House A
- Size
- 1000 sq. ft.
- Actual market price
- Hidden
Foundation Idea
Can an AI learn to predict the price of a house?
Start with one example, one input, and one prediction. Then watch how mistakes become feedback.
Learning
The process through which a model improves its predictions using examples from data.
One Neuron
Meet an artificial neuron
A neuron is a small adjustable formula. It receives an input, combines it with learned parameters, and produces an output.
Artificial Neuron
A basic computational unit that receives inputs, combines them using learned weights and a bias, and produces an output.
×
+
=
output
Select x, w, or b to see what it means.
Input
1000 sq. ft.
Weight
4.5
Neuron
Predicted Price
Rs. 55.0 lakh
Different weights and biases produce different predictions.
Prediction And Error
The network makes a prediction
Large error -> Large loss
Small error -> Small loss
Prediction Error
The difference between what the model predicted and what actually happened.
Loss
A numerical measure of how wrong the model's prediction is.
Rhetorical Pause
The model is wrong. How does it learn?
Feedback Loop
How a neural network learns
Training repeats the same cycle many times: predict, compare, send error information backward, adjust, and try again.
Training
The repeated process through which a model learns useful parameter values from data.
Backpropagation
A method for sending information about the prediction error backward through the network to determine what needs adjustment.
Optimizer
The mechanism that updates the model's weights and biases to reduce its error.
Training Example
Make Prediction
Compare With Actual
Calculate Error
Send Error Backward
Adjust Parameters
Watch Training
Watch the model learn
Press train to see repeated corrections. The prediction moves closer to the actual price and the error declines.
Learning does not mean storing one answer. It means adjusting parameters so that predictions improve across examples.
Round 1
Prediction: Rs. 55.0 lakh
Error: Rs. 5.0 lakh
Training Data
One example is not enough
The model needs many examples to detect useful patterns instead of reacting to one house.
800 sq. ft.
Rs. 45 lakh
1000 sq. ft.
Rs. 60 lakh
1500 sq. ft.
Rs. 80 lakh
1800 sq. ft.
Rs. 94 lakh
Training Data
Examples used by a model to learn patterns and adjust its parameters.
Many examples -> Detect patterns -> Adjust parameters -> Improve predictions
Zoom Out
From one neuron to a neural network
More inputs and more neurons let the model learn richer relationships, while the learning loop remains the same.
Neural Network
A collection of interconnected artificial neurons that learns relationships between inputs and outputs.
Hidden Layer
A layer of neurons that learns intermediate patterns from the input data.
Deep Learning
Learning with neural networks containing multiple layers that can discover increasingly complex patterns.
Size
Location
Bedrooms
Age
Predicted Price
Two Modes
Training is different from prediction
Training
- Training examples
- Prediction
- Error
- Adjust parameters
- Repeat
Inference
- New unseen house
- Trained neural network
- Predicted price
Using a trained model to make a prediction for new input.
Generalization
Has it learned or memorized?
Our model performs extremely well on the houses it studied. Is it necessarily a good model?
Training Data
Data used for learning.
OK OK OK OK OK
Test Data
New unseen data used to check whether learning generalizes.
OK X X OK X
Overfitting
When a model performs very well on its training examples but poorly on new unseen examples.
Good AI should learn patterns that generalize, not merely memorize examples.
Complete Story
The neural network learning story
Data
Predict
Compare
Measure Error
Learn
Adjust
Repeat
After Training
New Input -> Trained Neural Network -> Prediction
A neural network learns by making predictions, measuring its errors, and adjusting its parameters so that future predictions become better.