Deep Learning is an intersection of machine learning and artifical intelligence which is used to transform the world in various areas. Deep Learning automates intellingence using brain like algorithms (neural networks). One of the key parts of Deep Learning is that it makes us think about what it takes to be human and has the potential to automate a lot of skilled labor. This also means that there is a moral part to this new technology.

Fundamental Concepts

Deep Learning is a subset of methods for machine learning. Machine learning is dedicated to study/develop machines that can learn and is used to solve practical tasks. Machine learning mimics what we do and does things it wasn’t programmed for. It uses 2 main ways learning.

• Supervised Learning
• Unsupervised Learning

Supervised Learning can be used to take what we know to transform it into what we want to know. The mayority of works revolves around a classifier of some kind. Unsupervised Learning is often used to aid in the development of a Supervised Learning algorithm.

Unsupervised Learning also transforms data but it isn’t previously known what the output will be. An Unsupervised Learning model finds patterns in a data set clustering patterns of data which we can use to better understand our data and build a better Supervised Learning model.

Another way machine learning is split is through parametric and non-parametric models. Parametricism is about how the learning is stored and the method of learning.

• parametric model: fixed number of parameters
• non-parametric model: infinite number of parameters (data set)

Parametric models tend to use trial and error while non-parametric models tend to count and use probability. Supervised parametric learning is like a machine with a fixed number of knobs. When learning it fine tunes each knob. It does this by taking the data, making a prediction, observing the outcome and adjusting the knob. Unsupervised parametric learning uses a similar approach but focuses on grouping data.

Non-parametric learning is a class of algorithms where the number of parameters is based on the data. This means that most of the time it counts the data.