Whilst neural networks are not as complex as somebody starting out in Data Science might think, it would be wrong to say they are simple to learn. These deep learning networks will transform data until it can be classified into an output. It consists of neurons (for information processing, like the brain) which multiply an initial value by some weighting, works out a bias as new values come in and then normalises the output with a function. It is a bit like how our brain would try to develop concepts and learn about the environment but instead, based on mathematical functions and programming.
There are 6 types of neural networks used for different applications. The list below provides an overview of each. There is a lot more technical detail behind these available online and via publications to help work out which is best for your application.
In this neural network, the data or input travels in one direction which is why it is often known as the simplest. The applications tend toile in computer vision and speech recognition as where classifying the target classes are complicated. The feedforward network responds to noisy data and is quite easy to maintain.
- Radial Basis
This type of neural network considers the distance of a point in relation to the center. They are used for time-series calculations and system control amongst other applications. A radial basis network will use a set of prototypes along with other training examples and find the distance between and input and a prototype. The activation functions of artificial neurons drive outputs that can be represented in different ways to show how the network classifies data points.
- Multilayer Perceptron
Comprised of one or more layers of neutrons. Data is fed into an input layer and there may be one or more hidden layers providing levels of abstraction and predictions are made on the output layer which is also known as the visible layer. This is suitable for classification prediction problems where inputs can be assigned to a class or label. Data is commonly provided in tabular format like a CSV or Excel sheet.
Used for image classification, object detection and image segmentation. Networks have convolutional layers that act as hierarchical object extractors.
This type of neural network models sequences by applying the same set of weights recursively to the state of the aggregation at time (t) and input at time (t). The neural network is used for text classification texts, machine translation and language modeling.
A modular network is one composed of more than one neural network, connected by some intermediary. They can allow for more sophisticated use of basic neural network systems managed and handled in conjunction. Each individual network within the model should hope to accomplish some subtask of the wider objective.
We have provided a brief overview of popular machine learning algorithms and types of neural network. The technology and scripting behind them is of course far more technical than can be explained in a short article and does require previous experience in mathematics and data science fields.
However, for anybody beginning an AI journey, knowing about how these models function is important and it is worth getting to know them in more detail.