Deep learning is a sub-field of the study of machine learning. Deep learning and machine learning are similar in that they both create algorithms that can change themselves without direct user input. The difference between deep learning and machine learning is that deep learning makes use of artificial neural networks with many levels of neurons modeled after the human brain. Deep learning is done through training artificial neural networks with large amounts of data to create/predict outcomes based on a given input. One of the keys to deep learning is having a large amount of data to train the algorithm's artificial neural networks. Training the algorithm with a large data set with inputs and expected outputs allows the neural networks to develop patterns based on the inputs to accurately predict outcomes.
The most prevalent and successful task for deep learning is through supervised learning. Supervised learning is done through training neural networks with a set of inputs mapped to a set of outputs where there can be more than one input for each output. This allows deep learning algorithms to build neural networks with a defined set of predictors. The two main types of data used to train neural networks for deep learning algorithms are structured data and unstructured data.
Structured data is data that can found in a relational databases where categories, columns, and tables are defined with values. Structured data are usually text only, pre-defined data models, easy to search, and resides in relational databases as mentioned in the previous sentence. An example of structured data could be hours studied, time slept, and past grades which could ultimately be used to predict the success of a student on an exam. Training deep learning algorithms with defined input characteristics by using structured data creates a deep learning algorithm that is fast and has high performance. Deep learning algorithms built using structured data have proved successful when trained using a strong data set.
Unstructured data is not categorized like structured data but rather data that cannot be displayed in a database with rows and columns. Examples of unstructured data are image, audio, and video files. Deep learning algorithms trained with unstructured data can be used to recognize patterns in the data or to differentiate between parts of the data. When using unstructured data you must define values you are using, such as the values and location of pixels in a image. An example of unstructured data could be a set of animal pictures. This unstructured data could be used to train a deep learning algorithm and an artifical network to classify which picture contains which animal.
One of the key components of deep learning is neural networks. Neural Networks are built by the algorithm based on the given inputs and outputs. Neural networks comprise of many processing nodes that are interconnected and organized into layers. Data moves through these nodes. The nodes receives a data item in integer form and multiplies it by its associated weight, the the resulting numbers of a layer are added together to yield a single number.
Neural Networks build predictors based on the data by using ReLU functions or Rectified Linear Unit functions. Neural networks use ReLU functions to define correlations between different inputs in predicting the output.