AI Words That You Might Hear
A beginner’s glossary of AI Terms
Written by Matthew Pietrzykowski
Machine Learning
Machine Learning (ML) is a convenient way to describe classes of algorithms that are used to gain insight into data in a way that allows a certain amount self-instruction which, if properly designed & trained, achieves a robustness to changes in initial conditions that are lacking in other types of analytic methods.
Regression
Regression is a general term describing a model that explicitly defines a relationship between features of interest and a target. The term is most often used when the target is a continuous numeric dependent variable.
Deep Learning
Deep learning is a subset of ML approaches. A way to think of deep learning is like engaging in deep thought where one allows ideas to ruminate and interact with other possible outcomes in order to think through as many possible journeys that reach the same outcome to develop a more comprehensive representation of the data problem. In data science, the goal is to derive more “insightful” features from the input data based on the journey taken through an interactive network of relationships before culminating in the final output of the model.
Case-based Reasoning
Case-based reasoning is simply using previous, known solutions as a template for a solution to the current problem. The template will need to be adjusted to the new situation and tested, but once developed can be used in the situation being faced.
Neural Networks
Neural networks are a specific type of ML algorithm modeled after the human brain. The power from an artificial neural network comes from combining many simple computational units as a network where each of the simple units collaborate to achieve a complex, nonlinear model of the underlying data.
Self-Organizing maps
Self-Organizing Maps (SOMs) is an example of an unsupervised learning method or one that does not require a target. It is a method to uncover the hidden structure or patterns in a high dimensional data set by defining a lattice, grid, or 2D node structure in which each node is defined by different sets of weights that are tuned as the data set is iteratively examined. As the algorithm iterates over a complex data set the defined grid can be thought of as a frame upon which a picture is developed revealing the inherent structure of the the data. When unique regions are shown on the grid or map, those regions can be used to classify or segment the data.
NLP
NLP or Natural Language Processing describes the ways in which data scientists extract or transform free form text to a form that is more amenable to extracting insights and are commonly used in further downstream modeling.
Bayesian Belief Networks
A Bayesian Belief Network is a way to model the inter-dependencies of different features that employs a graphical approach to model the conditions that variables impose on other variables. The total set of graphical connections represent a network where the conditional dependencies are defined in terms of belief about the relationship governing that dependence.
Hidden Markov models
Hidden Markov models are a subset of Bayesian Belief networks where the conditional dependencies are modeled with an inherent sequential dependence, for example things happening in time. The fancy name just signifies that to gain a current understanding all one needs is information of the state one sequence step before.
Computer-vision
Computer vision from the perspective of machine learning is the process by which a device captures an image and processes the image to a form that can be used in common machine learning models. For example, automatically detecting a specific images of a horse from a series of pictures of similar animals, like cows, zebras, and other similar animals.
k-NN
K Nearest Neighbors are a set of algorithms that achieve a supervised outcome in an iterative manner that tests each case against a neighborhood in an effort to measure and assign each case to the neighborhood most similar to that case. The size of the neighborhood defining the similarity is given by k and set by the user. This approach is different from the traditional method of training a model with a full training instance space en masse.
Instance Based Learning
Instance based learning is an extension of the k-NN set of algorithms where the current instance informs the current state of understanding in an iterative manner, but without maintaining a “memory” of past states thus reducing size and time requirements that can be prohibitive.
Decision Trees
Decision trees are a class of algorithms the iteratively partition the input data space in a hierarchical way. So, let’s say a data set is split in half according to some criteria, say “dog” or “cat”. Now, each partition can be split further and further and further. These partitions can be represented as a tree, much like an organizational chart which helps represent the complex interactions in a palatable form.
Automation
Automation, in general, is the act of committing common tasks to a script, macro, or other series of steps that can be applied with minimal to no human intervention. With respect to machine learning, this means automating the process of model selection or other ML tasks. A common example is applying a grid search to model parameters in order to arrive an optimally tuned model.
Robotics
Robotics can be thought of coupling automation and engineering where an example would be using machine learning algorithms to learn defects on a production line and apply output of these algorithms to inform the signals used to control a mechanical arm that eliminates the defective product.
SVMs
Support Vector Machines (SVMs) are a class of algorithms that can be used in classification or regression where the governing ability is achieved by defining a smaller set of instances that truly define the decision boundary and optimizing that region. The points defining this region are supporting the boundary and can be set to be highly refined or more relaxed depending on the case. The definition of the actual boundary is very flexible and scalable by applying the math in one space but applied in higher order space.
Cognitive Learning
Cognitive learning extends the paradigm of machine learning to a more complex space encompassing the realm of human cognition to deal with much higher level decisions. This is the cutting edge and still being defined.
About the Author:
Matthew Pietrzykowski is a founding member of Humans For AI, a non-profit focused on building a more diverse workforce for the future leveraging AI technologies. Learn more about us and join us as we embark on this journey to make a difference!