The previous article gave a brief introduction to deep learning. This article deals with the networks used in deep learning. This network is known as a neural network. As the name suggests the network is made up of neurons
The networks used in artificial intelligence are a combination of blocks arranged in layers. These blocks are called an artificial neurons. They mimic the properties of a natural neuron. One of the neurons is the sigmoid neuron.
This is in general the formula for the sigmoid function. Every neural network consists of weights and biases.
Weights- The scalar quantities which get multiplied to the input
Biases- the threshold quantity above which a neuron fires
Working of a neuron
This is the simple representation of a neuron. This is similar to the biological neuron. In this neuron, the inputs are given along with some priority known as weights. The higher the value of the weights, the more prioritized is that input. This is the reason for our brain to choose one activity over the other. Activity is done only if the neuron fires. A similar situation is seen here. The particular activity is forwarded to the next layer only if this particular neuron fires. That is the output must be produced from the neuron.
Condition for the neuron to fire
The neuron will produce an output only if the inputs follow the condition.
As mentioned before, the bias is the threshold value and the neuron will fire only when the value crosses this bias. Thus the weighted sum for all the inputs must be greater than the bias in order to produce an output.
Classification of networks
Every neural network consists of three layers majorly: –
- Input layer
- Hidden layer
- Output layer
The input layer consists of inputs in the form of vectors. Images are converted into 1-D vectors. Input can be of any form like audio, text, video, image, etc. which get converted into vectors.
This is the layer in which all the computations occur. This is generally not visible to the user hence termed as a hidden layer. This layer may be single or multiple based on the complexity of the task to be performed. Each layer processes a part of the task and it is sent to the next layer. Vectors get multiplied with the weight matrix of correct dimensions and this vector gets passed onto the next layer.
The output layer gets information from the last layer of the hidden layer. This is the last stage in the network. This stage depends upon the task given by the user. The output will be a 1-D vector. In the case of classification, the vector will have a value high for a particular class. In the case of regression, the output vector will have numbers representing the answer to those questions posed by the user.
The next article is about the feed-forward neural network.