Artificial Intelligence is hot. We can hardly do anything without coming into contact, consciously or unconsciously, with forms of Artificial Intelligence. And it is becoming increasingly important. This article is an introduction to the field of Artificial Intelligence. It starts with a definition and then explores the different sub-specialties, complete with description and some applications.

WHAT IS ARTIFICIAL INTELLIGENCE?

Artificial Intelligence (AI) uses computers and machines to imitate people’s problem-solving and decision-making skills. One of the leading textbooks in the field of AI is Artificial Intelligence: A Modern Approach (link resides outside Axisto) by Stuart Russell and Peter Norvig. In it they elaborate four possible goals or definitions of AI.

Human approach:

• Systems that think like people
• Systems that behave like people

Rational approach:

• Systems that think rationally
• Systems that act rationally

Artificial intelligence plays a growing role in (I)IoT (Industrial) Internet of Things, among others), where (I)IoT platform software can provide integrated AI capabilities.

SUB-SPECIALTIES WITHIN ARTIFICIAL INTELLIGENCE

There are several subspecialties that belong to the domain of Artificial Intelligence. While there is some interdependence between many of these specialties, each has unique characteristics that contribute to the overarching theme of AI. The Intelligent Automation Network (link resides outside Axisto) distinguishes seven subspecialties, figure 1.

Each subspecialty is further explained below.

MACHINE LEARNING

Machine learning is the field that focuses on using data and algorithms to imitate the way humans learn using computers, without being explicitly programmed, while gradually improving accuracy. The article “Axisto – an introduction to Machine Learning” takes a closer look at this specialty.

MACHINE LEARNING AND PREDICTIVE ANALYTICS

Predictive analytics and machine learning go hand in hand. Predictive analytics encompasses a variety of statistical techniques, including machine learning algorithms. Statistical techniques analyse current and historical facts to make predictions about future or otherwise unknown events. These predictive analytics models can be trained over time to respond to new data.

The defining functional aspect of these engineering approaches is that predictive analytics provides a predictive score (a probability) for each “individual” (customer, employee, patient, product SKU, vehicle, part, machine, or other organisational unit) to determine, to inform or influence organisational processes involving large numbers of “individuals”. Applications can be found in, for example, marketing, credit risk assessment, fraud detection, manufacturing, healthcare and government activities, including law enforcement.

Unlike other Business Intelligence (BI) technologies, predictive analytics is forward-looking. Past events are used to anticipate the future. Often the unknown event is of significance in the future, but predictive analytics can be applied to any type of “unknown,” be it past, present, or future. For example, identifying suspects after a crime has been committed, or credit card fraud if it occurs. The core of predictive analytics is based on capturing relationships between explanatory variables and the predicted variables from past events, and exploiting them to predict the unknown outcome. Of course, the accuracy and usefulness of the results strongly depends on the level of data analysis and the quality of the assumptions.

Machine Learning and predictive analytics can make a significant contribution to any organisation, but implementation without thinking about how they fit into day-to-day operations will severely limit their ability to deliver relevant insights.

To extract value from predictive analytics and machine learning, it’s not just the architecture that needs to be in place to support these solutions. High-quality data must also be available to nurture them and help them learn. Data preparation and quality are important factors for predictive analytics. Input data can span multiple platforms and contain multiple big data sources. To be usable, they must be centralised, unified and in a coherent format.

To this end, organisations must develop a robust approach to monitor data governance and ensure that only high-quality data is captured and stored. Furthermore, existing processes need to be adapted to include predictive analytics and machine learning as this will enable organisations to improve efficiency at every point in the business. Finally, they need to know what problems they want to solve in order to determine the best and most appropriate model.

NATURAL LANGUAGE PROCESSING (NLP)

Natural language processing is the ability of a computer program to understand human language as it is spoken and written – also known as natural language. NLP is a way for computers to analyse and extract meaning from human language so that they can perform tasks such as translation, sentiment analysis, and speech recognition.

This is difficult, as it involves a lot of unstructured data. The style in which people speak and write (“tone of voice”) is unique to individuals and is constantly evolving to reflect popular language use. Understanding context is also a problem – something that requires semantic analysis from machine learning. Natural Language Understanding (NLU) is a branch of NLP and picks up these nuances through machine “reading understanding” rather than simply understanding the literal meanings. The purpose of NLP and NLU is to help computers understand human language well enough so that they can converse naturally.

All these functions get better the more we write, speak and talk to computers: they are constantly learning. A good example of this iterative learning is a feature like Google Translate that uses a system called Google Neural Machine Translation (GNMT). GNMT is a system that works with a large artificial neural network to translate more smoothly and accurately. Instead of translating one piece of text at a time, GNMT tries to translate entire sentences. Because it searches millions of examples, GNMT uses a broader context to derive the most relevant translation.

Learn how Natural Language Processing works (link resides outside Axisto).

Natural language processing – understanding people – is key to AI justifying its claim to intelligence. New deep learning models are constantly improving the performance of AI in Turing tests. Google’s Director of Engineering Ray Kurzweil predicts AIs will “reach human levels of intelligence by 2029“(link resides outside Axisto).

By the way, what people say is sometimes very different from what people do. Understanding human nature is by no means easy. More intelligent AIs expand the perspective of artificial consciousness, opening up a new field of philosophical and applied research.

SPEECH

Speech recognition is also known as automatic speech recognition (ASR), computer speech recognition or speech-to-text. It is a capability that uses natural language processing (NLP) to process human speech in a written format. Many mobile devices incorporate speech recognition into their systems to perform voice searches, e.g. Siri from Apple.

An important area of ​​speech in AI is speech-to-text, the process of converting audio and speech into written text. It can help visually or physically impaired users and can promote safety with hands-free operation. Speech-to-text tasks contain machine learning algorithms that learn from large datasets of human voice samples to arrive at adequate usability quality. Speech-to-text has value for businesses because it can help transcribe video or phone calls. Text-to-speech converts written text into audio that sounds like natural speech. These technologies can be used to help people with speech disorders. Polly from Amazon is an example of a technology that uses deep learning to synthesise human-sounding speech for the purposes of e-learning and telephony, for example.

Speech recognition is a task where speech is received by a system through a microphone and checked against a database of large pattern recognition vocabulary. When a word or phrase is recognised, it will respond with the corresponding verbal response or a specific task. Examples of speech recognition include Apple’s Siri, Amazon’s Alexa, Microsoft’s Cortana, and Google’s Google Assistant. These products must be able to recognise a user’s speech input and assign the correct speech output or action. Even more sophisticated are attempts to create speech based on brain waves for those who cannot speak or have lost the ability to speak.

EXPERT SYSTEMS

An expert system uses a knowledge base about its application domain and an inference engine to solve problems that normally require human intelligence. An interference engine is a part of the system that applies logical rules to the knowledge base to derive new information. Examples of expert systems include financial management, business planning, credit authorisation, computer installation design, and airline planning. For example, an expert traffic management system can help design smart cities by acting as a “human operator” to relay traffic feedback for appropriate routes.

A limitation of expert systems is that they lack the common sense people have, such as an understanding of the limits of their skills and how their recommendations fit into the bigger picture. They lack the self-awareness of people. Expert systems are not a substitute for decision makers because they lack human capabilities, but they can dramatically ease the human work required to solve a problem.

PLANNING SCHEDULING AND OPTIMALISATION

AI planning is the task of determining how a system can best achieve its goals. It is choosing sequential actions that have a high probability of changing the state of the environment incrementally in order to achieve a goal. These types of solutions are often complex. In dynamic environments with constant change, they require frequent trial-and-error iteration to fine-tune.

Planning is making schedules, or temporary assignments of activities to resources, taking into account goals and constraints. To design an algorithm, planning determines the sequence and timing of actions generated by the algorithm. These are typically performed by intelligent agents, autonomous robots and unmanned vehicles. When designed properly, they can solve organisational scheduling problems in a cost-effective way. Optimisation can be achieved by using one of the most popular ML and Deep Learning optimisation strategies: gradient descent. This is used to train a machine learning model by changing its parameters in an iterative way to minimise a particular function to the local minimum.

See also our “More Optimal Planning and Optimalisation Software”.

ROBOTICS

Intelligence is at one end of the Intelligent Automation spectrum, while Robotic Process Automation (RPA), software robots that mimic human actions, is at the other end. One is concerned with replicating how people think and learn, while the other is concerned with replicating how people do things. Robotics develops complex sensor-motor functions that enable machines to adapt to their environment. Robots can sense the environment using computer vision.

The main idea of ​​robotics is to make robots as autonomous as possible through learning. Despite not achieving human-like intelligence, there are still many successful examples of robots performing autonomous tasks such as carrying boxes, picking up and putting down objects. Some robots can learn decision making by associating an action with a desired outcome. Kismet, a robot at M.I.T.’s Artificial Intelligence Lab, learns to recognise both body language and voice and respond appropriately. This MIT video (link is outside Axisto) gives a good impression.

COMPUTER VISION

Computer vision is an area of ​​AI that trains computers to capture and interpret information from image and video data. By applying machine learning (ML) models to images, computers can classify and respond to objects, such as facial recognition to unlock a smartphone or approve intended actions. When computer vision is coupled with Deep Learning, it combines the best of both worlds: optimised performance combined with accuracy and versatility. Deep Learning offers IoT developers greater accuracy in object classification.

Machine vision goes one step further by combining computer vision algorithms with image registration systems to better control robots. An example of computer vision is a computer that can “see” a unique series of stripes on a universal product code and scan it and recognize it as a unique identifier. Optical Character Recognition (OCR) uses image recognition of letters to decipher paper printed records and/or handwriting, despite the wide variety of fonts and handwriting variations.

WHAT IS MACHINE LEARNING?

This article covers the introduction to machine learning and the directly related concepts.

Machine learning is the field of study that gives computers the ability to learn without being explicitly programmed. It is a subset of artificial intelligence (AI) and computer science that focuses on the use of data and algorithms to imitate the way humans learn, and in doing so it gradually improving its accuracy. By using statistical learning (link resides outside Axisto) and optimisation methods, computers can analyse datasets and identify patterns in the data. Machine learning techniques leverage data mining to identify historic trends to inform future models.

According to the University of California, Berkeley, the typical supervised machine learning algorithm consists of three main components:

• A decision process: A recipe of calculations or other steps that takes in the data and returns a guess at the kind of pattern in the data that the algorithm is looking to find.
• An error function: A method of measuring how good the guess was by comparing it to known examples (when they are available). Did the decision process get it right? If not, how do you quantify how bad the miss was?
• An updating or optimisation process: The algorithm looks at the miss and then updates how the decision process comes to the final decision so that the miss will not be as great the next time.

Machine learning is a key component in the growing field of data science. Using statistical methods, algorithms are trained to make classifications or predictions and uncover key insights from data.

HOW DOES A MACHINE LEARNING ALGORITHM LEARN?

The technology company Nvidia (link resides outside Axisto) distinguishes four learning models that are defined by the level of human intervention:

• Supervised learning: If you are learning a task under supervision, someone is with you, prompting you and judging whether you’re getting the right answer. Supervised learning is similar in that it uses a full set of labelled* data to train an algorithm.
• Unsupervised learning: In unsupervised learning, a deep learning model is handed a dataset without explicit instructions on what to do with it. The training dataset is a collection of examples without a specific desired outcome or correct answer. The neural network then attempts to automatically find structure in the data by extracting useful features and analysing its structure. It learns by looking for patterns.
• Semi-supervised learning: Semi-supervised learning is, for the most part, just what it sounds like: a training dataset with both labelled and unlabelled data. This method is particularly useful in situations where extracting relevant features from the data is difficult or where labelling examples is a time-intensive task for experts.
• Reinforcement learning: In this kind of machine learning, AI agents are trying to find the optimal way to accomplish a particular goal or improve the performance of a specific task. If the agent takes action that moves the outcome towards the goal, it receives a reward. To make its choices, the agent relies both on learnings from past feedback and on exploration of new tactics that may present a larger payoff. The overall aim is to predict the best next step that will earn the biggest final reward. Just as the best next move in a chess game may not help you eventually win the game, the best next move the agent can make may not result in the best final result. Instead, the agent considers the long-term strategy to maximise the cumulative reward. It is an iterative process: the more rounds of feedback, the better the agent’s strategy becomes. This technique is especially useful for training robots to make a series of decisions for tasks such as steering an autonomous vehicle or managing inventory in a warehouse.

* Fully labelled means that each example in the training dataset is tagged with the answer the algorithm should produce on its own. So a labelled dataset of flower images would tell the model which photos were of roses, daisies and daffodils. When shown a new image, the model compares it to the training examples to predict the correct label.

In all four learning models, the algorithm learns from datasets based on human rules or knowledge.

In the domain of artificial intelligence, you will come across the terms machine learning (ML), deep learning (DL) and neural networks (artificial neural networks – ANN). Artificial intelligence and machine learning are often used interchangeably, as are machine learning and deep learning. But, in fact, these terms are progressive subsets within the larger AI domain, as illustrated in Figure 1.

Therefore, when discussing machine learning, we must also consider deep learning and artificial neural networks.

THE DIFFERENCE BETWEEN MACHINE LEARNING AND DEEP LEARNING IS THE WAY AN ALGORITHM LEARNS

Unlike machine learning, deep learning does not require human intervention to process data. Deep learning automates much of the feature extraction piece of the process, eliminating some of the manual human intervention required, which means it can be used for larger data sets.

“Non-deep” machine learning is more dependent on human intervention for the learning process to happen because human experts must first determine the set of features so that the algorithm can understand the differences between data inputs, and this usually requires more structured data for the learning process.

“Deep” machine learning can leverage labelled datasets, also known as supervised learning, to inform its algorithm. However, it does not necessarily require a labelled dataset. It can ingest unstructured data in its raw form (e.g., text and images), and it can automatically determine the set of features that distinguishes between different categories of data. Figure 2 illustrates the difference between machine learning and deep learning.

Deep learning uses multiple layers to progressively extract higher-level features from the raw input. For example, in image processing, lower layers may identify edges, while higher layers may identify the concepts relevant to a human, such as digits or letters or faces.

In deep learning, each level learns to transform its input data into a slightly more abstract and composite representation. In an image-recognition application, the raw input may be a matrix of pixels. The first representational layer may abstract the pixels and encode edges; the second layer may compose and encode arrangements of edges; the third layer may encode a nose and eyes; and the fourth layer may recognise that the image contains a face. Importantly, a deep learning process can learn which features to optimally place in which level on its own. This does not fully eliminate the need for manual-tuning – for example, varying numbers of layers and layer sizes can provide different degrees of abstraction. The word “deep” in “deep learning” refers to the number of layers through which the data is transformed.

NEURAL NETWORKS

An artificial neural network (ANN) is a computer system designed to work by classifying information in the same way a human brain does, while still retaining the innate advantages they hold over us, such as speed, accuracy and lack of bias. For example, it can be taught to recognise images and classify these according to elements they contain. Essentially, it works on a system of probability – based on data fed to it, it can make statements, decisions or predictions with a degree of certainty. The addition of a feedback loop enables “learning” – by sensing or being told whether its decisions are right or wrong, it modifies the approach it takes in the future.

Artificial neural networks consist of a multilevel learning of detail or representations of data. Through these different layers, information passes from low-level parameters to higher-level parameters. These different levels correspond to various levels of data abstraction, leading to learning and recognition. An ANN is based on a collection of connected units called artificial neurons (analogous to biological neurons in a biological brain). Each connection (synapse) between neurons can transmit a signal from one neuron to another neuron. The receiving (postsynaptic) neuron can process the signal(s) and then signal to neurons connected to it downstream. Neurons may have state, generally represented by real numbers, typically between 0 and 1. Neurons and synapses may also have a weight that varies as learning proceeds, which can increase or decrease the strength of the signal that it sends downstream. Typically, neurons are organised in layers, as illustrated in Figure 3. Different layers can perform various kinds of transformations on their inputs. Signals travel from the first (input), to the last (output) layer, possibly after traversing the layers multiple times.

USES OF MACHINE LEARNING

There are many applications for machine learning; it is one of the three key elements of Intelligent Automation and a autonomous operating model within Industry 4.0. The computer programs can read text and work out whether the writer was making a complaint or offering congratulations. They can listen to a piece of music, decide whether it is likely to make someone happy or sad, and find other pieces of music to match the mood. In some cases, they can even compose their own music that either expresses the same themes or is likely to be appreciated by the admirers of the original piece.

Neural networks have been used on a variety of tasks, including computer vision, speech recognition, machine translation, social network filtering, playing board and video games, and medical diagnosis. As of 2017, neural networks typically have a few thousand to a few million units and millions of connections. Although this number is several orders of magnitude less than the number of neurons in a human brain, these networks can perform many tasks at a level beyond that of humans (e.g., recognising faces, playing “Go”).