WEBVTT

00:00.000 --> 00:03.800
Networks of neurons enable machines to learn from data.

00:03.800 --> 00:06.000
Let's learn neural networks.

00:10.200 --> 00:14.600
Artificial neurons, known as perceptrons, learn from data,

00:14.600 --> 00:18.800
while biological neurons process information naturally.

00:20.600 --> 00:24.400
In biology, information flows through neurons.

00:24.400 --> 00:28.600
Dendrites receive signals, the soma processes them,

00:29.000 --> 00:32.200
and the axon transmits the response onward.

00:33.800 --> 00:36.600
Artificial neurons work similarly,

00:36.600 --> 00:42.000
but instead of synapses, they use weights that control the strength of input signals.

00:43.200 --> 00:49.600
Understanding this biological inspiration helps us to see why neural networks are so powerful.

00:50.800 --> 00:54.600
Every neural network begins with the basic perceptron,

00:54.600 --> 00:59.600
the simplest unit that combines inputs, multiplies them by weights,

00:59.600 --> 01:03.600
adds a bias, and applies an activation function.

01:03.600 --> 01:09.600
This simple structure lets networks learn by adjusting weights like neurons do.

01:09.600 --> 01:17.600
This idea helps solve real-world problems like filtering emails and recognizing handwriting.

01:18.200 --> 01:21.200
Neural networks grow from simple ideas,

01:21.200 --> 01:26.600
starting with perceptrons and expanding into multi-layer networks.

01:26.600 --> 01:32.600
It makes decisions using a linear boundary, suitable for data separable by a straight line.

01:34.000 --> 01:37.200
However, perceptrons have limitations.

01:37.200 --> 01:42.600
They cannot solve problems that require curved or complex decision boundaries.

01:42.600 --> 01:47.400
Perceptrons are limited, so we use multi-layer networks.

01:47.400 --> 01:52.400
The input layer receives the data, hidden layers handle most of the learning,

01:52.400 --> 01:56.400
and the output layer produces the predictions.

01:57.400 --> 02:02.400
During a forward pass, each input is multiplied by its corresponding weight,

02:02.400 --> 02:07.400
summed with a bias, and then passed through an activation function.

02:07.400 --> 02:13.400
Common activation functions include sigmoid, tang, and relu.

02:13.400 --> 02:19.400
Each helping the network introduce non-linearity, so it can model complex patterns.

02:19.400 --> 02:25.400
Although neural networks can look complex, their core computation is surprisingly simple.

02:25.400 --> 02:35.400
From the image, key points include inputs such as 0.5-0.2, enter the neuron.

02:35.400 --> 02:42.400
Each input has a corresponding weight such as 0.8-0.4.

02:43.400 --> 02:49.400
A bias, for example, 0.1, shifts the output.

02:49.400 --> 02:57.400
They combine using a weighted sum equals np. inputs comma weights plus bias.

02:57.400 --> 03:02.400
Then an activation function relu is applies.

03:02.400 --> 03:06.400
Output equals max of 0 and weighted sum.

03:06.400 --> 03:10.400
Though simple, this shows how artificial neurons work

03:10.400 --> 03:16.400
and helps learners gain confidence before exploring larger neural networks.

03:16.400 --> 03:23.400
Mastering these basics prepares students for advanced machine learning and deep learning.