But what is a neural network? | Deep learning chapter 1
A visual construction of a small digit-recognition network: pixels become activations, layers transform them, and weights and biases determine the result. The video motivates what ‘learning’ changes, while deliberately leaving the training algorithm to the next chapter.
Plan for this page
Watch actively with preparation, signposts, and a comprehension check.
You will leave with
What layers, activations, weights, biases, and activation functions do in a simple feed-forward network.
Time
19 min
Before you begin
Curious beginners who are comfortable following pictures and a small amount of algebra.
Do this now
Read the terms and attention points before loading the video.
But what is a neural network? | Deep learning chapter 1What layers, activations, weights, biases, and activation functions do in a simple feed-forward network.
00:04 · The digit-recognition problem
02:43 · Neurons, activations, and layers
05:32 · A hope for layered features
08:21 · Weights encode influence
10:14 · Sigmoid and bias
What it teaches
What layers, activations, weights, biases, and activation functions do in a simple feed-forward network.
Curious beginners who are comfortable following pictures and a small amount of algebra.
Pay attention to
Treat the drawn neurons as numbers, not miniature biological brains.
Follow how one layer’s values produce the next through weighted sums.
Notice that the hoped-for edge and loop detectors are an intuition, not a guarantee about what the trained network represents.
Useful terms
Before you press play
activationhidden layerweightbiasactivation function
The external video loads only when you choose
The player uses the official privacy-enhanced address. The provider’s terms still apply.
A neuron in this model is a numerical activation; the brain analogy is loose.
Weights and biases are learned parameters controlling how input patterns influence later layers.
Learning is the process of finding useful parameter values; this video explains the structure, not the full training process.
Active check
In the network shown, what does ‘learning’ mainly change?
Caveats
The 2017 network is intentionally small and plain; convolutional networks, transformers, and other modern architectures differ substantially.
The visual story that layers discover neat human-readable edges and loops is motivating intuition. Trained representations can be distributed or unlike the concepts a person expected.
Sigmoid is used for exposition; the closing segment explicitly notes the prevalence of ReLU in deeper networks.
Accessibility
Manually supplied English and French caption tracks and an English transcript were available at review. The mathematical animation is central: narration is strong, but learners who cannot see the diagrams may need the official text lesson or a described alternative.
inAi