A-Z of AI Jargon Busting
If you’re looking for a particular word or phrase and know what letter it starts with, select it below to start searching:
Jargon Index
A, B, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, U, VA
Adversarial Algorithms
What are they?
If there is any term that is likely to conjure images of killer robots, it’s the idea of adversarial algorithms although these are actually techniques which we benefit from every day. The way the technique works is by trying different permutations of input data in order to force the algorithm to incorrectly classify the record, so as to discover potential vulnerabilities.
What are they good for?
These techniques are commonly used in spam filtering applications, and other cyber security use cases.
Anomaly Detection
What is it?
This is a technique to discover outliers in a particular data set. The method can be tuned in order to pick out examples that differ from the norm to a greater or lesser extent, although in the case of supervised anomaly detection, a human expert would first flag examples of anomalies and a feedback method would be used to continue the training on real data sets.
What is it good for?
The method can be used in analytic contexts such as fraud detection, where no single attribute might cause concern, but the pattern of behaviour when taken as a whole doesn’t fit the ‘normal’ data set.
Apriori Algorithms
What are they?
These are algorithms that are used to scan databases in order to build rules that can be used to make assumptions about the association between data in the dataset.
What are they good for?
This method is useful in terms of performing basket analysis in a retail context. Although it has limitations in terms of the computational resources that it requires to train, the resultant rules tend to be fairly simple in runtime to perform rudimentary next-best action or recommendations.
Associative Rules
What are they?
Associative rules are the product of techniques such as Apriori Algorithms. The rules are in the form of; in X% cases where there is Y, then Z will be also present.
What are they good for?
Although expensive to generate, they are simple to run and can be used for recommendations. They are also easy to understand and often cited in the context of X-AI (explainable AI).
B
Back-propagation
What is it?
The term ‘back-propagation’ is short for “backward propagation of errors” and is a method originally devised in the 1970s as a solution to doing differentiation of complex nested functions in an automated manner.
Until this point, this type of mathematics would need to be hard-coded or done by hand – which was time consuming and error prone. As multi-layer neural networks needed training to be effective, the discovery of back-propagation as an alternative (and better) method to discover features in the mid-1980s was a key breakthrough that has led to modern AI.
What is it good for?
As a method for optimisation, back-propagation is very efficient – particularly on modern GPU architecture which is often tuned for this task. The method however is good at discovering local maxima/ minima as opposed to global maxima/ minima.
C
Capsule Neural Network
What is it?
A Capsule Neural Network is a special form of convolutional neural network (CNN – see below) that is intended to better resemble the way neurons are organised in nature.
What is it good for?
It is particularly useful in situations where hierarchy is important, and to defeat adversarial methods where, for example a small change to an image causes it to be incorrectly classified.
Classification Analysis
What is it?
Classification, along with clustering is a primary technique in AI applications. It is simply the method to group items with have similar attributes into categories. Classification requires supervision – but once items have been classified (such as images that contain street signs) and the model trained, features of these images that were used to train the model are used to identify further examples that contain similar features.
What is it good for?
Classification is good to sort items into their particular types. For example, a classification engine might be used to separate employment contracts from employment application forms.
Clustering Analysis
What is it?
Along with classification methods, clustering is a primary method used in AI applications. Clustering simply separates examples into groups which exhibit similar characteristics.
What is it good for?
Clustering tends to be used in workflows where supervision isn’t possible or practical. It is useful in determining families of data which are related and can be then subject to further manual analysis.
ChatBot
What is it?
A ChatBot is an application that is able to enter into a dialogue with a human by parsing natural language questions to elicit intent, and thereby responding appropriately with either the required answer or follow-up questions that can guide the chatbot to being able to ultimately provide the user with the answer they are seeking.
What is it good for?
Chatbots are powerful in cases where the answer cannot simply be found using a search engine, or in cases where a workflow needs to be initiated through the text interface (such as opening a new account, or changing an address).
Cognitive Computing
What is it?
Cognitive computing is a term that is used to describe applications of AI where multiple methods are linked together or operated in parallel in order to solve a particular problem.
What is it good for?
Cognitive computing is not a technique or method in itself, but is simply a term that is used to describe a particular solution implying a level of complexity that is similar to that which a human would need to exhibit. The exemplar of this method is the Jeopardy! playing AI system that IBM developed, which needed to consider multiple paths of solving the problems in order to determine the correct answer.
Collaborative Filtering
What is it?
Collaborative filtering is an approach that assumes that where a common value exists between two records, the value of a third variable is likely to be informative as to a missing value in the data set.
What is it good for?
It is a method that is used in recommendation systems. If a user has expressed their preference for a number of examples, then their preference can be inferred based on other users who have expressed similar preferences.
Computer Vision
What is it?
Computer Vision is the general term applied to how to perceive and understand images or videos by computers.
What is it good for?
It’s application is as diverse as industrial processes (machine vision) to image enhancement and restoration. Security systems such as facial recognition are built on components of the wider field of computer vision.
Conversation Engine
What is it?
A conversation engine is simply a piece of software that is able to extract intent from natural language input (either typed or spoken), and compose responses that mimic the responses that might be given by a human.
What is it good for?
Conversation Engines are powerful tools in automation efforts. They can be used to simply replace the information gathering tasks otherwise performed by a human customer service agent, leaving the human expert to focus their attention on dealing with the matter in hand – or in advanced cases can even take place of the human agent altogether.
The question as to whether the conversation engine ought mimic humans entirely or even disclose their true nature is an active topic within the field of digital ethics.
Convolutional Neural Networks
What are they?
Convolutional Neural Networks (CNNs) are the archetypal ‘deep learning’ method that is prevalent today. It is a modification of the older method of multilayer perceptrons that passes values forward between layers in order to distinguish features in a data set.
What are they good for?
CNNs are an example of a mathematical model for performing AI that is inspired by our understanding of natural biological processes. In the case of CNNs, they closely resemble the process taking place within the area of the brain responsible for visual tasks, the so-called ‘visual cortex’. They are typically used in image and video recognition applications.
D
Decision Trees
What are they?
A decision tree is a visual representation of the permutations of a particular operation or process. They typically resemble the root system of a tree, the entry point of the process being the stem, which subdivide into major arteries and further branches at the various decision points along the process.
What are they good for?
Historically they have been used as a tool in operational effectiveness design, but latterly they have been used in machine learning to optimise strategies in order to achieve a particular goal (such as the minimum number of questions to ask a patient in order to achieve a correct diagnosis).
Deep Belief Networks
What are they?
A Deep Belief Network is a form of hierarchical machine learning method where layers are generated in order to analysis data but with no connections between units in each layer.
What are they good for?
DBNs are particularly effective at optimisation problems such as creating trading strategies or in medical applications such as drug discovery.
Deep Learning
What is it?
Deep Learning is a set of methods used in AI applications that have the common characteristic of analysing data for patterns based on multiple layers of processing where the result of analysis from one layer is passed to the next. The term originated in the mid-1980s, although appears to have taken on a mystique in the current age as a catch-all term for all effective machine learning methods.
What is it good for?
As a general term it is hard to be specific about where the strengths and weaknesses lie, but it’s fair to say that most pattern recognition processes have benefited from the introduction of deep learning approaches in recent years (i.e. speech recognition, text recognition etc).
E
Eclat Algorithms
What is it?
Similar to Apriori algorithms, this is a method to scan a dataset in order to build rules to support prediction of future values.
What is it good for?
The Eclat method is very fast, and less computationally intensive than other methods – although the memory requirements for using such a method might prove the limiting factor. As with Apriori algorithms, these are often used as the basis of a recommendation engine.
Evolutionary Algorithms
What are they?
Evolutionary Algorithms are simply types of algorithms that are designed to approximate mechanisms that are seen in nature such as mutation, reproduction, and selection.
What are they good for?
Evolutionary algorithms tend to be very good at optimisation problems, but can be very expensive in terms of the computational requirements necessary to execute them. Often a seemingly simple algorithm can create an almost intractable level of computational complexity.
Expert Systems
What are they?
An expert system is a piece of software that takes the place of a human expert in decision making through reasoning between rules that are pre-programmed, often on instruction of the human expert. They originated in the 1960s, but became very popular in the early 1980s in the second-AI bubble; although are now no longer typically referred to as an AI method due to their inability to ‘learn’ independently of their programming. Expert systems comprise two separate subsystems. The Knowledge Base which contains facts and rules, and the Inference Engine which is the logic that enables decisions to be reached.
What are they good for?
The textbook on Expert Systems by Hayes-Roth was published in 1983. It lists 10 applications of expert systems which are; interpretation, prediction, diagnosis, design, planning, monitoring, debugging, repair, instruction, and control. Expert systems carry one great advantage over other types of AI in that the decisions reached are easily explainable. The disadvantage is that once a certain level of complexity in the rules is reached, often inconsistencies arise that the inference engine cannot make sense of. If this does not become the limiting factor, then computational requirements often does. Finally, if these two limitations aren’t reached – the inability to ‘learn’ from new data limits their practical application to simple knowledge based applications.
F
Factor Analysis
What is it?
Factor Analysis is a method aimed at modelling a smaller number of variables in data to those observed.
What is it good for?
As with PCA and LDA, Factor Analysis is used in Financial Services, Marketing and Bioscience – but it’s best known application is in psychometrics.
Feature Extraction
What is it?
Feature extraction is a technique to reduce the complexity of data by grouping or combining elements into features for subsequent processing.
What is it good for?
It’s a powerful technique that enables stronger interpretation of the data, for example in an image – large areas of differently coloured green pixels might be grouped and a human would interpret that ‘grass’ as a feature as been extracted. It’s also used as a method in Optical Character Recognition systems (OCR) where the outline of letters is extracted and compared to known letter shapes.
Feedforward Neural Network
What are they?
A feedforward neural network is a simple and alternative method to RNNs where connections between nodes do not form a cycle. Its simplest implementation is the single-layer perceptron where the results of the input nodes are weighted and fed directly to the output nodes with no hidden layer. They were the earliest form of machine learning that was devised, as early as the 1940s were theories of how natural processes operate (particularly hearing and sight) translated into what now are considered rudimentary mathematical forms.
Later, multi-layer perceptrons were developed which can solve a much larger number of problems and thus are a major component of modern AI research and design.
What are they good for?
Feedforward neural networks can be used for the same problems that other techniques such as RNNs or CNNs might be used for, but with a greater chance that they will ‘overfit’ the model to the training data – and thus require more data to train to a serviceable degree of accuracy.
Owing to their relative simplicity they are mostly confined to elementary explanations of neural network theory before the practitioner is introduced to the more complex (and more powerful) CNN and RNN methods.
Forecasting and Prediction
What is it?
These are not AI techniques as much as they are the consequent objectives from AI. The two terms are not interchangeable – although often colloquially are so. Forecasting is a special case of prediction which time is the operative vector. A forecast is a prediction of future values in a time series, whereas a prediction can simply be of missing values in a data set (whether future, past, or non-time series).
What are they good for?
If one thinks of AI as being either about methods to interpret the world, or to predict the future – then its clear that these are two key aspects. Depending on the application – humans are good at both forecasting and prediction – given sufficient data, so to now are machines.
FP-Growth Algorithms
What are they?
Frequent Pattern Growth Algorithms are another example of associative rules. The count of items above a certain threshold in the dataset is recorded, after which the FP-tree structure is built by inserting instances.
What is it good for?
It’s a fast method of doing basket analysis or used as part of a recommendation engine.
G
Gaussian Mixture
What is it?
A mixture model is a method for representing categories of data within the entire dataset. It’s a probabilistic method, so the accuracy of identifying subpopulations depends on the appropriate model selection as well as the quality and volume of data available. Gaussian mixtures are one class of such models, and can be Bayesian, non-Bayesian or multivariate.
What is it good for?
Disaggregation is a typical use case for such a method. Imagine your electricity supplier receives a once a second data transmission with the current meter reading of your house. They would like to know which model fridge and television is being used. Even though there is no meta-data being transmitted with the signal, the ‘fingerprint’ of each device can be ascertained with a level of probability. Once the devices being used are identified, predictions for future energy consumption (and therefore energy markets trading) are much more accurate.
Generative Adversarial Networks
What is it?
Generative Adversarial Networks (GANs) are an application of deep neural networks where two networks are created and pitted against each other. The first, called the ‘generator’ creates new data instances, while the second ‘discriminator’ compares them against real-world examples.
What is it good for?
GANs have been used for a number of different applications, but the most famous of which are to create AI versions of artwork in a particular style, or indeed music that seems like from a certain composer. In fact, GANs are so powerful – that the results can fool the lay-audience into believing the output is the real thing.
Genetic Algorithms
What is it?
Genetic Algorithms are a type of evolutionary algorithm that are inspired by the existence of chromosomes and genotypes in nature. These are mutated and recombined through generations of algorithms in order to find a solution that is better fitted for the problem.
What is it good for?
Genetic Algorithms have found their application in engineering applications, such as designing the optimal shape for wind resistance. Often the resultant design appears to the eye as counter-intuitive, but can be deceivingly effective.
Geospatial Analysis
What is it?
This is simply the technique of using data relating to the physical world (such as mapping data, or geological survey data) in order to predict the existence of certain features.
What is it good for?
In the Oil & Gas industry, this technique is used to predict the potential location of oil underground or undersea.
H
Hierarchical Temporal Memory
What is it?
While Deep Learning captures all the public imagination, within the research field – Hierarchical Temporal Memory (HTM) as an alternative method of machine learning are gaining significant ground. The technique is based on a theory of natural intelligence originally published in 2004 by Jeff Hawkins and Sandra Blakeslee in their book On Intelligence.
HTMs can be thought of multiple layers of algorithms that ‘learn’ from the data by looking at different aspects of the data and storing certain results that relate to attributes of time or space.
What is it good for?
HTMs can be used as an alternative method to solving the problems tackled by deep learning approaches. While still a relatively nascent field – commercial application of HTMs have found success in anomaly detection, classification, as well as language processing and generation.
I
Image Analysis
What is it?
Image Analysis is simply the goal of taking a digital image and extracting actionable information from the data.
What is it good for?
Early examples of this would be barcode scanners, now QR code readers are the latest such variant. Advanced applications are automatic image captioning. Facial recognition systems are derivatives of this approach.
Image Processing
What is it?
Image Processing is the technique of applying transformations to some or part of an image in order to facilitate further processing or analysis.
What is it good for?
Instagram is an entire business built on the principles of image processing. A digital image is transformed into something which has a higher subjective appeal using filters (a particular type of image processing technique.
Inference
What is it?
Building an AI system typically involves first ‘training’ it using data before it can be deployed to solve a problem. The problem is that training a model is massively computationally intensive, and while this is a necessary cost of designing something that is fit for purpose (or simply good enough), one doesn’t want such a high cost of compute in operation.
Inference is the second step, a form of optimisation that reduces the complexity of the model to something which can get the same result without the computational overhead.
What is it good for?
Almost all applications of deep learning use inference in run-time. It simply gets the same result as the ‘trained’ model, but without the necessary infrastructure to support it.
Insights Visualisation
What is it?
This term has a lot less to do with AI and much more to do with data analytics. Any way of presenting data can be considered visualisation, with the exception of presenting information in tables.
What is it good for?
Insights visualisation is simply highlighting in a visual way the aspects of the data which are important to the user.
K
k-Means Clustering
What is it?
Breaking a data set down into the categories that it contains is the job of clustering. The number of clusters is denoted by the letter ‘k’. The standard approach is to randomly select k of the total data (n) and then each other data point is categorised as being one of the k clusters based on which of the originally labelled points it is closest to. Once this has completed, the centre of each cluster becomes the new mean and the process is repeated until the clusters don’t change.
What is it good for?
The technique was originally used in radio transmission as a way to send analogue signals in a digital format. It’s now used for a wide variety of usages – often a way of sorting documents into their various types (i.e. employment contracts, policies, and certificates would require k=3).
Keyword Analysis
What is it?
Keyword analysis is the technique of taking a corpus of text and representing the key words that are present in it, or comparing the text to ascertain how closely or far it matches a given set of keywords.
What is it good for?
Keyword analysis has its roots in digital marketing. It’s the cornerstone of internet searching. What you enter into a search engine such as Google is compared against keywords from millions (billions even) of websites, and then based on this (and now other factors such as how many other pages link to the page in question, and the authority of the page) the results are returned to you in ranked order.
L
Linear Regression
What is it?
It’s a straight line through data plotted on a chart with a linear scale.
What is it good for?
Linear regression is either used a tool to make predictions on data or as an aid to explain variation from the norm.
Linear Discriminant Analysis
What is it?
Linear Discriminant Analysis (LDA) or Fisher’s Method, named after British statistician, Sir Ronald Fisher; is a method used for classification of data.
It’s related to PCA methods and Factor Analysis in that they all look for linear features to explain the data. LDA and Factor Analysis models the difference between the classes of data whereas PCA models the similarities.
What is it good for?
LDA has many applications, and is used in industries as diverse as biomedical, marketing, and financial services. It’s most widespread application is in facial recognition systems.
Logistic Regression
What is it?
Logistic Regression (or Logit Modelling) is a form of regression analysis where the dependent variable is binary.
What is it good for?
It’s particularly useful to model data against a probability of an event occurring (i.e. the likelihood of a driver having an accident in their first year of driving against their age on passing the driving test).
Long Short-Term Memory
What is it?
Long Short-Term Memory (LSTM) is a form of RNN which overcomes some technical limitations known as ‘vanishing gradients’ under certain circumstances. They still suffer from another problem caused by training RNNs using back-propagation which is that of ‘exploding gradients’.
A gradient is simply an expression of a derivative as a vector. Vanishing gradients are those which tend to zero, and exploding gradients are those which tend to infinity.
What is it good for?
LSTMs are often used in robotics, and also applications where forecasting (time-series prediction) is a key component.
M
Machine Vision
What is it?
Machine Vision is the application of computer vision technology to industrial processes.
What is it good for?
Machine Vision is used in manufacturing processes, as well as drone/ robot guidance. It is one of the primary technologies behind autonomous vehicles (along with battery/ motor and guidance systems).
Monitoring & Alerts
What is it?
This is the application of AI to the stream of log data created by many IoT and other complex systems in order to ascertain anomalies within the data feed that might indicate reliability issues or malevolent attack.
Alerting capability might either filter noise from a human decision-maker, or indeed notify another system which would then make decisions based on the data feed towards some pre-defined goal or goals.
What is it good for?
Cyber security, prescriptive analytics, and dynamic maintenance systems.
Multilayer Perceptrons
What are they?
A multi-layer perceptron is one where at least one hidden layer is in existence that weights the results of the input layer once again before passing the result to the output layer. They are typically feedforward in nature, although can also be convolutional or recurrent.
They are typically trained using the back-propagation method.
What are they good for?
Multi-layer perceptrons were once the prevalent method of machine learning and in the 1980s were used to advance areas of computer science such as speech and image recognition.
They are now predominantly used as a way of training people in machine learning techniques owing to their relative simplicity.
N
Naïve Bayes
What are they?
Naïve Bayes are a family of algorithms that are used as a simple method for constructing classification engines. It’s a statistical method that compares the likelihood of given assumptions being true based on input rules. They are typically trained in a supervised context.
What are they good for?
Naïve Bayes classifiers are surprisingly good even at seemingly complex problems, which considering the advantage of the necessity for a limited training data set is impressive.
They have fallen somewhat out of fashion in recent years, particularly since the mid-1990s with the rise of random forests.
Natural Language Generation
What is it?
Natural Language Generation (NLG) is one of the few areas in the field of AI where the label describes exactly what the field is about.
It is a subset of Natural Language Processing techniques, and can be thought of as distinct from template driven systems (where bodies of text or numbers are inserted into the corpus based on input data) as in the case of NLG the language produced is dynamic and not following predefined rules.
There is a grey area between complex template systems (such as those which might produce legal documents) and true-NLG systems.
What is it good for?
NLG has only in very recent times found commercial application, but areas as diverse as autonomous commentary (of charts or data tables), sports reporting, and chatbots all use these systems.
To demonstrate the level of quality that NLG can reach, the book Stories from 2045 published in 2019 contained three short stories that were generated using a HTM based NLG model.
Natural Language Processing
What is it?
Natural Language Processing (NLP) is the field of taking human language and enabling a machine to ‘understand’ it in a way such as a human might.
The most famous example of NLP is the so-called ‘Turing test’ based on an article published by Alan Turing in 1950 which described the point where a machine’s responses are indistinguishable from a human’s responses being the cross-over point in artificial intelligence research. Turing envisaged this point would be reached by about the year 2000. At time of writing, the Turing test has not yet been exceeded, and consensus is that we are at least a decade away from achieving this breakthrough.
What is it good for?
Another famous implementation of NLP was the ELIZA programme (named after Eliza Doolittle from the book Pygmalion) created in 1966 by Joseph Weizenbaum. The idea behind this programme was to mimic the type of responses a Rogerian psychotherapist might offer. While the intention of the programme was to demonstrate the superficiality of machine generated responses, Wiezenbaum’s secretary appeared to develop affection towards the programme which became a pivoting moment in his career after which he became one of the leading proponents on limits to computer science.
NLP technology is today ubiquitous in chatbots and digital assistants such as Siri, Cortana and Alexa. It can also be combined with other techniques in order to build sophisticated fraud detection systems as well as a tool to build advanced recommendation engines.
Nearest Neighbour
What is it?
The nearest neighbour algorithm is a solution to the travelling salesman problem. The problem is “what is the shortest possible route that visits each city in a list of cities and returns to the start, if the distance between each pair of cities is known”.
It is one of a class of problem known as non-deterministic polynomial-time hardness (NP-hard). In simple terms this means that as the number of variables increases (in the case of the travelling salesman problem, this is the number of cities) so the problem becomes exponentially more difficult.
What is it good for?
The nearest neighbour method picks a random city and simply visit the nearest unvisited city until all have been visited. It’s advantage is that it is simple to understand and usually quick to compute, but rarely does it find the optimal solution.
Neural Networks
What is it?
A neural network (or more precisely, an artificial neural network) is a computational model where the outputs of nodes (computational units) are transmitted to other nodes for further computation in a way that somewhat mimics the natural processes that we understand to occur in animal brains.
What is it good for?
Neural networks are ubiquitous today, and the leading method of developing artificial intelligence systems.
One should be mindful however of Moravec Paradox which states that it’s relatively easy for a computer to perform complex tasks that an adult human can achieve, yet disproportionately difficult to achieve simple tasks (such as motor function, balance, and perception) that even a young infant can master.
Neural networks are the cornerstone of breakthroughs in machines tackling complex tasks, and are slowly making inroads into the ‘simpler’ human abilities albeit in very narrow areas.
P
Perceptron
What is it?
A perceptron is the building block of modern supervised machine learning. It is an algorithm that classifies its inputs based on weightings.
It was originally imagined to be a system that would be physically implemented. It was invented by American psychologist Frank Rosenblatt in 1957 who tried to replicate with electronics what he observed in nature.
A machine was indeed built, ‘the Mark 1 perception’ but the method was first encoded in software as a demonstrator. It is curious how most of AI research has been implemented in software, and it is only in recent years that again custom hardware is being built to solve these problems.
Rosenblatt gave a demonstration of the perceptron in 1958 where he very enthusiastically talked-up the potential for his invention. This led the New York Times to report on the discovery as “… the embryo of an electronic computer that… will be able to walk, talk, see write, reproduce itself and be conscious of its existence”.
This overhype in the media is perhaps what led Marvin Minsky and Seymour Papert to labour the limitations of perceptrons in their 1969 book Perceptrons which was widely mis-cited and is often blamed for the first AI winter. Minsky and Papert were later criticised as they were already aware of solutions to the limitations of single-layer perceptrons in multi-layer neural networks.
What is it good for?
In itself, not much – but in combination with other perceptrons; particularly in deep multi-layer networks – then with sufficient data and computational resources, models can be trained to exceed human levels of ability, albeit in narrow instances.
Principal Component Analysis
What is it?
Principal Component Analysis (PCA) is a form of Factor Analysis, but looks at total variance in the data before transforming the original variables into a smaller set or ‘principal components’ for further processing.
What is it good for?
PCA has a number of applications outside of AI, such as in quantitative finance where the complexities of modelling a significant number of financial instruments can be reduced to a small number of factors which bear a high correlation to easily modelled indices. It can be used as a method for stock picking or indeed for risk modelling.
As with common factor analysis, PCA also has application in neuroscience, and it is from this discipline that it is applied as a tool for data analytics and predictive modelling. It is the simplest method in the class of analytics known as ‘eigenvector multivariate analysis’ and often is thought of as the simplest way of explaining the internal structure of data.
Profile & Trend Analysis
What is it?
Profile and trend analysis is a simple method of comparing the gradient between multiple data series.
What is it good for?
It’s a method to quickly assess whether there might be correlation between data series.
Q
Q-Learning
What is it?
Q-Learning is a form of reinforcement learning developed by Christopher Watkins for his PhD thesis in 1989. He was inspired by how animals could be conditioned to learn by giving them a reward, and so sought to recreate this in an algorithm that could use a similar approach to learning.
What is it good for?
Q-Learning is a powerful technique and a variation of it was patented by Google DeepMind in 2014 which was able to play early computer games at an equivalent level to human experts.
R
Random Forests
What is it?
A problem with decision trees is that they overfit to their training data, and so a solution to this is randomising decision trees in high quantities, and then taking the average from the set.
What is it good for?
It’s a highly effective method of classification and regression, and is very commonly used because of its simplicity and versatility.
Recurrent Neural Network
What is it?
Recurrent Neural Networks (RNNs) are a type of neural network where state can be passed to process a sequence of inputs.
What is it good for?
RNNs are often used in cases where context is important. For example, in OCR technology – the probability of a given letter appearing after another is dependent on the earlier letter and the word that it sits in (and indeed, the likelihood of a given word appearing after another is depending on the immediately preceding word). The letters ‘o’ and ‘u’ might look similar in a low quality scan of a document, but if they appear immediately after a ‘q’, it is highly unlikely that the ‘o’ is present, and therefore even if the letter looks like an ‘o’, it is really likely to be a ‘u’.
Reinforcement Learning
What is it?
Reinforcement learning is a method of training a model based on a given reward (or punishment) for a given sequence of tasks or steps. It’s an entirely separate method of training to supervised or unsupervised methods, and typically requires brute-force computation to achieve results.
What is it good for?
Reinforcement learning is often used to train industrial devices to navigate complex environments, and is often done through use of simulated environments until the system reaches a sufficient level of competence. Autonomous vehicle navigation systems are often an implementation of Reinforcement Learning.
The famous examples of Google DeepMind playing the ancient Chinese game Go and beating human experts is reinforcement learning par excellence.
Roboadvisor
What is it?
A robo-advisor a special case implementation of a chatbot that asks a series of questions about an individual’s appetite to risk or financial goals, and then constructs a financial planning strategy that would mimic that of a human financial advisor.
The term is a colloquialism that breaks the convention of physical manifestations of AI being referred to as robots and software implementations being called bots. Correctly, roboadvisors should be instead named advisorbots.
What is it good for?
As an alternative to human financial advisors, roboadvisors gained significant popularity in the aftermath of the 2008 global financial crises. The quality of the investment decisions made by such systems is the subject of frequent controversy and speculation, but the compelling argument for their continued development and existence is the fact that they are able to address an underserved market. It was previously unaffordable for those making less than six figure investments to hire professional financial advice, whereas the economies of roboadvisory services mean that the same service provision can now be offered to those merely making three figure investments.
S
Sentiment Analysis
What is it?
Sentiment Analysis is a subset of NLP where words, phrases, and paragraphs are analysed for their sentiment or conformity to a set of base criteria.
What is it good for?
Sentiment Analysis can be used as a simple method for aggregating the verbosity of human data. For example, rather than asking users to rate products using a five-star method, they could instead be asked to give a review of the product in their own words and the sentiment analysis engine be used to generate the star-rating.
Such a method would likely lead to much lesser arbitrariness in the ratings awarded (such systems usually have a disproportionately high numbers of top, bottom or middle awards due to the apathy of the user).
Simulation based Optimisation
What is it?
Optimisation is the process of tuning a model for accuracy or computational efficiency, or a combination of these two factors.
Simulation based Optimisation is the method of achieving optimisation by varying parameters in a simulator.
What is it good for?
Simulation based Optimisation can potentially dramatically speed up the training of a model, particularly where the source input data is sparse. Applications for this method exist in controlled process modelling such as industrial environments where edge-cases need to be discovered that might not be explicitly considered by the designer.
Social Network Analysis
What is it?
Social networks are a representation of relationships between people based on a particular factor. In the case of Facebook, it is the presence of a mutual ‘friendship’ indicator, content ‘likes’ and other interactions. In the case of an Enterprise environment, it is the communication content, latency and frequency that can be used to infer the social network.
What is it good for?
The analysis of social networks can be used for a number of applications, the most common of which is digital marketing or fraud detection. A criticism of social network analysis is that their insights can potentially create feedback loops that self-amplify.
Speech Analysis
What is it?
There are two aspects to speech analysis technology, speech recognition and emotion detection.
The first, which has been developed since the early 1950s is simply the task of taking the human spoken word and transforming it into text for further processing. Early attempts were limited to numbers and a small number of predefined keywords trained to a single individual voice. Development in this field has been to increase the speaker independence, vocabulary (and language) completeness, and tending to real time recognition. Amazon’s Alexa and Apple’s Siri are at the pinnacle of this 70 year endeavour.
The latter technique is to recognition the emotion in the speakers voice in order to detect non-verbal communications such as sarcasm or wit where the intent of the speaker might be contrary to the words used and thus the sentiment derived. A similar process can also be used to detect they betrayal of unintentional verbal clues such as deception that might be present through hesitation or uncharacteristic modulation of the voice.
What is it good for?
The goal of speech analysis has always been to foster more natural communication between humans and machines. Whereas most of the interface between humans and machines has been until recently physical input (keyboard, mouse, touchscreen) and visual output (indicator lights and screen), we’re currently reaching a tipping point where speech analysis and voice synthesis technology is reaching a level where some tasks are more efficient to be driven by voice command than by physical control or display output (asking if it will rain today, or voice security authentication are but two such examples).
Support Vector Machines
What are they?
Support Vector Machines (SVMs) are a powerful method that is commonly used to do data classification. They are supervised learning models that were originally developed in the 1960s but the implementation used today ‘soft-margin’ SVMs were proposed in the mid-1990s.
What are they good for?
SVMs are widely used, but their most common application is in image analysis such as classification of images or text recognition.
T
Temporal Difference
What is it?
Temporal Difference (TD) is a method of reinforcement learning that measures the difference between the predicted reward and the actual reward in order to adjust the model.
What is it good for?
This method compares predicted results against actual results and therefore is good for situations where predictions are being made of data that will soon be known such as weather forecasting. The long range forecast might be inaccurate, but will continuously improve as previous long range forecasts are compared to the actual readings.
It is a computationally intensive process, but one which appears to mimic natural processes within animal brains. Behavioural science is an area of research that is advancing the understanding of TD models.
Temporal Relation Identification
What is it?
In process modelling, when analysing a stream of data collected from inputs such as cursor movements or mouse clicks – certain activities can be identified as discrete courses of action and connected together using temporal relation identification.
What is it good for?
The data might predict that the user is going to copy and paste data from a website contact page into a CRM. The individual activities which make up this process are discrete and are order dependent (i.e. the data cannot be copy and pasted until the webpage has fully loaded).
Time Series Analysis
What is it?
When data is sampled and date/time stamped then it can be represented as a time-series. This might be a random polling (i.e. date/time stamps of speed camera images) or ‘discrete-time’ data which occurs at a regular interval (i.e. at temperature at midday in Trafalgar Square).
What is it good for?
Analysis of time-series data is useful to discover trends that are hidden within the data. A famous example from an Israeli parole board found that convicts were more likely to be granted parole if their hearing took place immediately following a lunch break. Another example is where a Bank instituted a policy that workers who stayed later than 8pm could order food deliveries to the office at company expense, and those who stayed after 9pm could expense taxi transport home. Time series analysis demonstrated a high instance of expense fraud as there was a significant drop in food orders between 8.30pm and 9.10pm, and then a spike of correlated food delivery orders and taxi bookings between 9.10pm and 9.30pm.
Topic Extraction
What is it?
Topic extraction is a subset of NLP where words can be grouped into clusters or ‘topics’ in order to ascertain the weighting of any particular keyword to the corpus of text within which it lives.
What is it good for?
Topic extraction is very useful the context of document management systems. Previous methods required human users to manually tag the documents, but the level of subjectivity in the meta data often hid the true meaning. With modern methods, regardless of how a document is tagged or labelled, it’s true meaning can be likely ascertained. Searching therefore for documents that pertain to ‘employee rights’ is therefore much more accurate using Topic extraction methods than previously possible.
V
Video Analysis
What is it?
Video analysis is a particular instance of image analysis that factors the complexity of a moving image.
What is it good for?
Video analysis can be used for a number of uses, particularly security applications where suspicious activity can be flagged by the model across a much wider set of captured data than human security officers would otherwise be able to achieve.