Layers and Deep Learning
Follow simple features as they become complete objects.
Layers and Deep Learning explained visually
Follow pixels through hidden layers until simple clues become a complete object.
Three roles in a network
The input layer receives original data, hidden layers create intermediate representations, and the output layer produces the final prediction. Hidden does not mean mysterious; it means we supply neither their features nor their answers directly.
Early layers find simple patterns
In an image network, early layers may respond to edges, brightness changes, and colors. Later layers combine those clues into curves, textures, eyes, ears, and other meaningful parts.
Depth builds a hierarchy
Every layer builds on the representation before it. Multiple hidden layers create a deep neural network, allowing a difficult recognition task to be divided into manageable stages.
More is not automatically better
A deeper network still needs suitable data, careful training, and an appropriate design. Extra layers require more computation and can learn unhelpful details when the training process is poorly controlled.
Quick check
What does the input layer receive?
The original numerical data, such as image pixels.
Why are layers called hidden?
Their intermediate features are not directly supplied as training answers.
What makes a neural network deep?
It contains multiple hidden layers.
Lesson 4 transcript
Imagine identifying a face in a photograph. You do not begin by naming the person from one pixel. Your visual system combines many small clues. It notices edges, curves, eyes, and larger facial shapes before recognizing the complete face.
A neural network can organize information in layers. The input layer receives the original data. For an image, this data comes from pixel values. The output layer produces the final result, such as dog, zebra, or horse.
Between them are hidden layers. We call them hidden because we provide the input and observe the final answer, but we do not directly supply the intermediate features that these layers create.
An early hidden layer can respond to simple patterns such as vertical edges, color changes, or bright areas. A later layer can combine edges into curves and shapes. Deeper layers can combine those shapes into ears, eyes, stripes, or complete object parts.
This creates a hierarchy of features. Each layer receives information from the layer before it and builds a more useful representation. The network can learn these features from data instead of requiring a person to design every visual rule manually.
Information usually moves forward from the input toward the output when the network makes a prediction. We call this a feedforward neural network. A network with multiple hidden layers is commonly called a deep neural network, and using these networks is part of deep learning.
More layers do not automatically guarantee better results. The network still needs suitable data, an appropriate design, and effective training. A very large network can also require more computing power and may learn unhelpful details if the training process is poorly controlled.
The important idea is that layers divide a difficult task into useful stages. Simple patterns appear first. Later layers combine them into more complete concepts. The output layer turns those concepts into a prediction.
In the next lesson, we will see how feedback changes the weights and allows the entire network to learn.