All language subtitles for 001 Understanding Deep Learning

In this section, we will go through the fundamental concepts of deep learning.

To understand deep learning, we have to know its position among similar or related fields.

Artificial intelligence is the biggest field among all.

The term artificial intelligence is made up of the words artificial and intelligence.

Artificial refers to human creations, and intelligence refers to the capacity for understanding.

AI enables machines to comprehend similarly to humans.

Machine learning is a subset of artificial intelligence.

One of technique in machine learning is neural networks.

Deep learning is a kind of neural networks which has many distinct layers and give a more sophisticated

results than simple neural networks.

Now we are going to review a fundamental concept of neural networks.

It is well known that neural networks can solve challenging problems by mimicking the behavior of the

human brain.

The input data is processed through several stack layers of artificial neurons to create the desired

output.

What is a neuron anyway?

The neuron, also known as a nodal unit, is the basic unit of computation in a neural networks.

After receiving input from another node or from an external source, the note processes the input and

produces output.

Its input has an associated weight.

W the function F is applied to the weight.

It's somebody.

No.

Function F is called activation function.

This whole unit is called the perceptron, the simplest neural network architecture.

Here are the various components of a perceptron.

Input is the group of features that the model used to learn.

For instance, an array of pixel values from an image can be the input for object detection.

With its primary function is to prioritize those features that contribute the most to learning.

Bias.

Its role is to see the value produced by the activation function to the left or right to fit the prediction

with the data of better.

The activation function's purpose is to introduce nonlinearity into the neurons output.

This is crucial as a lot of real world data is non-linear, and we want the neural networks to learn

this nonlinear data.

The perceptron represents how a single neuron performs.

What if we take a lot of perceptron?

We will have a multi layer networks with the input data pass in the forward direction.

The layer of the feedforward neural networks contains several neurons or nodes.

Notes in adjacent layers have connections, and each connection has a weight w.

Feedforward networks can consist of three types of nodes.

Input nodes bring data from the outside world into the networks.

No computation is performed at the input node.

Only information is passed to the hidden node.

He did not computes and transfers data from the input not to the output node.

Computation will be performed and networks output will be generated to put No.

In general, deep learning networks have numerous hidden layers which make them called deep neural networks.

We could not find any solid reference on how many hidden layers that make a neural networks called deep

neural networks.

However, a deep learning hidden measures are associated with extracting features, whereas standard

neural networks use neurons to transmit input to get output with the help of various connections.

Now we are going to talk about why the residual network or Internet is very important.

More hidden layers typically give networks the chance to learn more effectively.

However, it could lead to gradient problems like fencing or exploding.

The product of any derivatives will occur in a network within hidden layers.

The gradient will decrease exponentially if the derivatives are small, which causes the vanishing gradient

problem.

However, if the derivatives are significant, the gradient will grow exponentially, creating the exploding

gradient problem.

If anything gradient occurs, the gradient may be zero and the network training will stop.

The moderates may grow very large if an exploding gradient occurs leading to overflow or nan, which

prevents the weight from being updated any longer.

For those issues.

Rest nets offer an alternative solution.

The fundamental principle of risk net is to copy the prior result x and end to the subsequent result

f x.

The previous slides explain how perceptron performs.

There is a function F in every perceptron, which is a nonlinear function known as an activation function.

So what is exactly an activation function?

An activation function is inspired by brain activities in which different neurons are activated by different

stimulus.

For example, when we touch file, certain neurons are activated, causing us to believe that we have

thus far feel the pain of the heat and immediately remove our hand.

Inactivation function in neural networks makes the choice of whether to activate a neuron or not.

Therefore, the activation function is a straightforward mathematical calculation to decide whether

or not the input from the neuron is significant during the prediction process.

Aside from that, the purpose of an activation function is to add nonlinearity to the neural networks

without activation functions and neural networks is simply a linear regression model that can struggle

to learn any complex task.

There are several activation functions available.

The sigmoid activation function will take a real valued input and put it between zero and one.

The hyperbolic tangent activation function will take a real valued input and push the value into the

range of values minus one and one.

Rectified linear unit, which is abbreviated.

This review is one of the most popular activation functions.

It's a very straightforward formula.

But both accuracy and speed are performed well.

This activation function only limits the number of zero.

This means that if X is less than zero, then if x equals zero and if x is greater than zero, then

if x equals x.

Miki will modifies review by allowing small negative values when values are less than zero.

Lakeville is used in all of three.

This is a self regularized non monotonic activation function when compared to review.

The missed function consistently improves the accuracy of the neural networks architecture.

This activation function is used in all of four.

Serial stands for sigmoid Weighted linear unit.

This activation function is calculated by multiplying the sigmoid function by its input.

Cielo is used in your office.

Seven.

See you in the next video.