Recently I have been invited to write some material about Artificial Intelligence, Machine Learning and Deep Learning. However there are already several good resources about these topics, so in the interest of knowledge transfer and to make your learning curve smoother I instead selected, filtered and summarized relevant information that you should know in this (short) blog post.
Let’s start with some basic definitions . Artificial Intelligence is an umbrella term for any program which can sense, reason, react and adapt. Machine Learning refers to algorithms which improve over time and more data (ML is an subset of AI). Deep Learning is subset of ML which specializes in multilayered neural networks which can learn from vast amounts of data. There are different ML problems, including Supervised learning, Unsupervised learning and Re-inforcement learning.
Training a model in ML means to iteratively improve it to minimize erros, maximize accuracy and feed data to the ML algorithm to maneuver on unfamiliar terrain.
Machine Learning tasks and problems can be classified depending on the output type, for example , in (statistical) classification inputs (a.k.a. events) are grouped into classes or labels (e.g.: spam filters) whereas in regression (analysis) the outputs are continuous because the numeric dependency is predicted / approximated (e.g. price of a house). Classification and Regression problems are of the Supervised learning category.
There are different types of artificial neural networks , including: Multi-Layer Perception (MLP), Convolutional Networks (CNN) and Recurrent Networks (RNN). There some sub-types of Recurrent Networks (vanilla RNN), including Long-Short Term Memory (LSTM) and Gated Recurrent Units (GRU), amongst others. A practical example of MLP could be an algorithm that takes as an input an image of some handwriting (e.g.: numbers) and translates it to its digital representation.
Artificial neural networks are made of neurons, connections and weights and the learning rule/algorithm. The learning process involves the selection/definition of a cost function and the calculation of the gradient (a gradient is a vector-valued function) of the loss function (backpropagation – a loss function is a function of a value of one or more variables with a real number to represent the cost).
There are several current and future applications of AI/ML/DL, including content customization (social media), recommender systems for retail (incl. property recommenders for real state), intelligent bots (incl. customer service and conversational commerce), gaming (to anticipate moves in a play), sports performance and Investment Portfolio Management (a.k.a.: https://en.wikipedia.org/wiki/Algorithmic_trading).
Frameworks and Tools
There are multiple tools and frameworks, being the following Python open source libraries the most popular options: TensorFlow (by Google), Caffe (by UC Berkeley) and Theano (by Université de Montréal) among others. Intel has optimized these popular libraries to run much faster on Intel architecture , more specifically for Intel Xeon (codename Broadwell) and Intel Xeon Phi (codename Knights Landing).
Microsoft also offers platforms (incl. Services such as Cognitive Services, Bot Framework and Azure Machine Learning and infrastructure such as Azure Cosmos DB), solutions for end users (e.g.: Dynamics 365 AI-based solution for customer service) and frameworks (e.g. Visual Studio Code Tools for AI , Microsoft Cognitive Toolkit and Cortana Skills Kit) for Machine Learning. There are also some benchmarks  suggesting that CNTK is better than TensorFlow.
Javier Andrés Cáceres Alvis