Understanding Artificial Intelligence - Major concepts for enterprise applications

1. UNDERSTANDING ARTIFICIAL INTELLIGENCE WHITEPAPER MAJOR CONCEPTS FOR ENTERPRISE APPLICATIONS
2. OVERVIEW ARTIFICIAL INTELLIGENCE - A TOPIC WE SHOULD DISCUSS WHAT IS AI? Where does AI come from, where does it go? SEGMENTATION Elements of AI, what they are and what they do TERMS & CONCEPTS Buzzwords that are important to understand 2 3 5 13 About this whitepaper Artificial Intelligence is a fundamental topic – for us as humans, as a society but also for businesses. For business executives and decision-makers it is sometimes hard to keep up with rapidly evolving technologies as part of the day-to-day business. By providing this curated compilation of information about the fundamental aspects of AI, we want to captivate and inspire you to become more involved with the technology by better understanding the underlying concepts and value drivers of this technology. WHERE CAN TECHNOLOGY TAKE YOU? Appanion in a nutshell We provide high-quality insights on real-world technology-driven business applications and help you as innovation partner with actionable strategies, technology knowhow and business model competence. I hope you can take away some valuable insights and concepts to improve your business! Tobias Bohnhoff Founder
3. 3 WHAT IS AI? DEFINING THE SPACE
4. DEFINING THE SPACE: WHAT IS ARTIFICIAL INTELLIGENCE? Artificial Intelligence (AI) is the ability of a computer-controlled entity to perform cognitive tasks and react flexibly to its environment in order to maximize the probability of achieving a particular goal. The system can learn from experience data, and can mimic behaviors associated to humans, but does therefore not necessarily use methods that are biologically observable. Sources: McKinsey, Stanford University, Wikipedia, European Commission 4 „ What are we talking about? When talking about Artificial Intelligence (AI) it is important to distinguish between systems designed to solve problems in a very specific context – the so called Narrow AI – and a human-like Artificial General Intelligence (AGI). As of today, AGI is a future scenario and requires a rather philosophical approach to be discussed in depth. Some experts say it is decades away to build such an AI, some argue that human will never be able to build something at an equal level of intelligence. Most often, AGI scenarios outline a dystopia for the human race due to the uncontrollable effects this technology potentially has. Narrow AI on the other hand is already automating or augmenting many tasks in a business context or even in the private life. Self-driving forklifts in warehouses, Amazon product recommendations, real-time trades on stock exchanges, interacting with Siri, person recognition in the iPhone photo album – everything is already done, without the need of human intervention. Shortcomings of most definitions Existing definitions of Artificial Intelligence usually rely on two terms which require precise definitions themselves: „Machine“ and „intelligence“. The extremely broad range of appearances of AI technology in the form of apps, physical robots or simply as cognitive functions (e.g. facial recognition) makes it almost impossible to describe it exhaustively. Our approach Due to the lack of a predominantly accepted definition, we decided to blend various definitions from McKinsey, Stanford University, the European Commission and Wikipedia in order to derive a satisfying basis:
5. 5 SEGMENTATION ARTIFICIAL INTELLIGENCE 5 SEGMENTATION ARTIFICIAL INTELLIGENCE
6. STRUCTURING THE DIFFERENT AI DOMAINS IS CHALLENGING AS EVERYTHING IS CONNECTED 6 Cognitive domains There are a lot of buzzwords around AI and also a lot of confusion about how all these different terms and application areas fit together. At first, it is important to distinguish between the „where“ and the „how“. Where means: In which part of intelligent cognitive functions an artificial algorithm is working or can be applied. This includes: ▪ Contextual environment perception ▪ Learning and concluding from experience ▪ Interpretation and processing of natural languages ▪ Representation of knowledge ▪ Planning of future actions ▪ Autonomous processing of actions based on environmental input data (robotics) How this can be done involves different techniques that are most often inseparable from an area of application. Application sub-domains Sub-domains of AI can be mapped to different cognitive functions but they can also occur in applications that are mainly focusing on other functions. To give an example: reinforcement learning can also be applied as a part of a robotics application but in this case the app combines both, the cognitive functions of autonomous action processing and learning. Another common connection exists between the perception of audio through machine hearing and speech recognition as part of natural language processing for example in smart voice assistant applications. Applications are most often composed of different cognitive functions and consequently combine their corresponding subdomains. Cognitive domains Application sub-domains AIAIAI Natural Language Processing Machine learning Machine Perception Intelligent Robots Automated Planning Knowledge Representation & Reasoning Computer vision Text Speech Software assistants Physical Motion and manipulation Planning Scheduling Ontologies Logic and probability theory Unsupervised Learning Supervised Learning Reinforcement Learning Computer audition Machine touch
7. MACHINE LEARNING ENABLES DECISION MARKING BEYOND HUMAN CAPABILITIES 7 Machine Learning is the science of programming computers so they can learn from data and/or information, and improve their learning autonomously. DEFINITION ORIENTATION Machine Learning focuses on teaching computers how to learn without the need to be programmed for specific tasks. It is commonly divided into the major sub-domains of supervised and unsupervised learning. In addition, there are two hybrid forms. In supervised learning input and output variables are provided, whereas in unsupervised learning only the input data is given. Choosing to use either a supervised or unsupervised machine learning algorithm typically depends on factors related to the structure and volume of your data and the use case of the issue at hand. Supervised learning is used in cases with labeled data like image or speech recognition, forecasting and the training of neural networks Unsupervised learning is used with unknown, unlabeled data during exploratory analysis or to pre- process data Semi-supervised learning uses both methods at the same time, when labeling it too time intensive or extraction of data difficult Reinforcement Learning iteratively uses punishment and reward for any outcoming result and learns from experience to train an algorithm „ “ Classification & Categorization Clustering Reduction of Dimensionality Regression Supervised Unsupervised DiscreteContinuous Prediction Detection Optimization APPLICATION
8. NATURAL LANGUAGE PROCESSING ENABLES FRICTIONLESS COMMUNICATION WITH COMPUTERS 8 Natural language processing (NLP) is a sub-dimension of artificial intelligence that deals with the written and verbal communication between computers and humans in natural languages. NLP is used to recognize, understand and interpret natural language input data and generate contextually meaningful output data. DEFINITION ORIENTATION Speech recognition describes methodologies and technologies that allow recognition of natural language out of recorded audio data and translation into text. Natural language understanding generally requires four major aspects: ▪ Lexicon (vocabulary and meaning of words) ▪ Parser (syntax analysis and structuring) ▪ Grammar (structural rules for the composition of words) ▪ Semantic theory (derivation of a logical, contextual meaning) Natural language generation systems translate data into language by making decisions on how to choose and compose the right words in order to effectively deliver a message to a human recipient. ▪ Content determination = WHAT information is included ▪ Document structuring = WHERE to place content parts ▪ Aggregation = MERGING of redundancies ▪ Lexical choice = HOW to formulate meaningful ▪ Expressions = WHICH references make contextually sense ▪ Realization = CREATION sticking to syntax, grammar, orthography Text analytics Text generation Automated communication „ “ APPLICATION
9. ROBOTS AUGMENT OR SUBSTITUTE HUMAN LABOR TO IMPROVE EFFICIENCY, SECURITY AND PRECISION 9 Robotics is a technology branch that aims to develop machines that can replicate human actions - partly even with enhanced capabilities DEFINITION ORIENTATION Robots capture environmental data via sensors and actuators and are able to perform tasks or a series of actions autonomously or semi-autonomously. Due to the extremely broad range of appearances, a clear definition of 'robots' is tricky. Usually, a robot is considered as any device that: ▪ is programmable ▪ performs (complex series of) tasks (semi-) autonomously ▪ interacts with the (physical or virtual) world based on its captured environment data Robotics and AI are connected, but neither exactly the same nor is Robotics simply a subset. The intersection between AI and robotics is the field of intelligent robotic systems, which operate not solely based on a set of pre-programmed rules. Intelligent Robots make choices based on their input data, react flexibly to their environment and learn from experience. Three major factors allow the intelligent robot segment to grow rapidly: ▪ Increasing sensor density and ability to capture precise real-time data ▪ Increasing miniaturization and performance of computational power ▪ Increasing mobility of devices thanks to ubiquitous wireless connectivity Service robots Autonomous vehicles Industrial robots Humanoid robots „ “ APPLICATION
10. KNOWLEDGE REPRESENTATION AND REASONING IS THE BASIS FOR CONTEXT CONSIDERING AI APPS 10 Knowledge representation and reasoning describes a subset of artificial intelligence that deals with the representation of information in a way that a computer is able to process it correctly. The goal is to solve complex tasks which require commonly accepted facts and rules to be applied in this context. DEFINITION ORIENTATION Knowledge Representation and Reasoning is... 1. A surrogate, used to determine consequences in advance by thinking rather than acting 2. A set of ontological commitments - setting a perceptive focus or selecting imperfect representations of the world, which determines our way of thinking 3. A theory of intelligent reasoning, dealing with the questions of what are defining properties of intelligent reasoning, what can be inferred at all and what should be inferred 4. A medium for pragmatically efficient computation - the guidance for organizing information so as to facilitate making the recommended inferences 5. A medium of human expression - the language in which we say things about the world Intelligent information representation Automation tasks that require continuous reasoning „ “ What is knowledge? Taking the world to be one way and not another What is representation? Symbolic encoding of propo- sitions believed by an agent What is reasoning? Producing logically derived representations of new propositions. The earth is spherical = true The earth is flat = false = elephant = avocado = David is Sue’s father = David is an adult male APPLICATION
11. Major areas of machine perception are computer vision, computer audition and machine touch, whereas tasting and smelling are in a very early stage of experimental research. Machine perception is valuable for many different areas of AI. It enables – or at least significantly reduces the effort of – data capturing and pre-processing for further applications such as natural language processing, robotics or machine learning. Computer vision mimics and automates the human visual system. It includes all tasks of acquiring, processing and analyzing digital images or videos in order to derive an understanding of the environment. Computer audition deals with the understanding of audio signals apart from natural language (see NLP). It includes representation, reasoning and grouping of audio signals based on general sound semantics. Machine touch processes haptic or tactile information that arise from the physical interaction with the environment. Re-sampling, filtering, channeling, noise reduction etc. Lines, edges, interest points, texture motion, auditory or surface modeling etc. Pattern recognition and selection of relevant sequences or samples Multi-input channel verification, object size and characteristics, classification Match / no-match, pass / fail or flagging for further inspection Data acquisition Pre-processing Feature extraction Detection High-level processing Decision-making WITHOUT MACHINE PERCEPTION ANY AI APPLICATION WOULD BE LITERALLY SENSELESS 11 Machine perception encompasses methods and technologies to simulate human senses and therefore the perception and subconscious interpretation of the environment. DEFINITION ORIENTATION Computer vision Computer audition Machine touch „ “ APPLICATION
12. AUTOMATED PLANNING SETS THE TONE FOR AUTONOMOUSLY ACTING INTELLIGENT MACHINES 12 The Planning problem in Artificial Intelligence is about the decision making performed by intelligent agents when trying to achieve a particular goal. It involves choosing a sequence of actions that will most likely transform the state of the world to satisfy that goal. DEFINITION ORIENTATION A plan is a structure that gives appropriate actions to apply in order to achieve an objective when starting from a given state. Why do we need Automated Planning at all? It is an important part of rational intelligent behavior and therefore a required element of building autonomous acting intelligent machines and software. Planning: Decide WHAT actions to use to achieve (a set of) goals Scheduling: Decide WHEN and HOW to perform a given set of actions The basic structure of an Automated Planning process includes three important elements: ▪ Planner ▪ Controller ▪ State transition system Depending on time and resource restrictions, it is possible to further include a scheduler that triggers re-planning, if the objective cannot be achieved due to the restrictions. Workflow Control Robots Videogames „ “ APPLICATION
13. 13 TERMS & CONCEPTS WHAT YOU NEED TO KNOW
14. ARTIFICIAL NEURAL NETWORKS ARE THE BASIS FOR MANY MAJOR ADVANCES IN MACHINE LEARNING Artificial Neural Networks (ANN) are frameworks inspired by biological brains to process complex data inputs by using various machine learning algorithms. Neural networks in general don’t need to be programmed but learn from training. Artificial Neural Networks work based on connected artificial neurons (nodes) that process signals by receiving input, changing their state (activation) and transmitting output. The transmission depends on the activation, input signals are pooled to one single output type, which is then forwarded to other connected neurons. Structure of artificial neural networks: ANNs are organized in layers. A data signal is processed from the input layer through neurons and edges (connections) of one or multiple hidden layers to an output layer. 14 Artificial Neural Networks Each neuron receives weighted inputs from the edges and computes an activation and output function to create a single output signal. Sometimes a threshold is used to limit the activation of neurons. The signal is further processed to the next connected neurons and so forth until the output layer is reached. The capabilities of the network are defined by the weighted graph of connections between the neurons. In some cases a bias term is added to the total weighted sum of inputs to influence the activation function. Network architecture (layers and machine learning algorithms) should be designed to meet the desired goal function (e.g. image recognition). In order to achieve a particular goal, fine-tuning in form of training (learning) is required. This can be done by modification of connections, weights and thresholds or even activation and output functions. ANNs use learning algorithms to automatically improve their inherent structural settings to produce the favored output. Input layer (image data) Output layer (object recognition) Hidden layers (shapes, textures etc.) Weighted edgesArtificial neurons (activation function) f(x) f(x) f(x) f(x) f(x) f(x) f(x) f(x) f(x) f(x) f(x) f(x) Activation function weighting threshold check multiple input signals single output aggregation function computing
15. AT A GLANCE: FURTHER IMPORTANT AI TERMS TO UNDERSTAND Algorithms are unambiguous specifications or rules to follow in order to solve a problem. Almost everything can be reduced to an algorithmic function. Algorithms perform calculation, data processing and reasoning tasks. Models in the context of artificial intelligence describe systems that use mathematical concepts and language such as statistics, game theory, logic etc. Models are usually composed of variables and equations to describe relationships between those variables. This happens in the form of mathematical functions. 15 Supervised learning describes a machine learning model design that is suitable for categorization, classification and regression analysis (determination of the relation between variables). The method requires labeled input data. The labeling process normally has to happen upfront and oftentimes requires manual effort. With a training and validation data subset, an algorithm is then trained to produce sophisticated output results. Another testing data subset finally assesses the model fit. For every supervised learning task the output characteristics are known beforehand. Unsupervised learning describes a machine learning model design that is suitable for clustering and reduction of dimensionality. The method works with unlabeled and often dynamically changing input data. The model learns relationships between elements of the input data by itself by searching for patterns. Deep Learning is a machine learning method using artificial neural networks with multiple hidden layers as its core framework to process data and predict outputs based on the goal function.
16. 16 LET’S TALK ABOUT BOOSTING YOUR BUSINESS WITH ARTIFICIAL INTELLIGENCE! We identify suitable AI use cases and help with target-oriented prioritization. IDENTIFY We create a concrete plan about requirements, project organization and feedback-driven iteration steps. PLAN We evaluate suitable solutions for your company-specific requirements EVALUATE We facilitate the development and implementation process and support you in the selection of partners. EXECUTE We provide decision confidence through process design and functional prototypes TEST We develop a future-proof data strategy based on the company's long-term goals SUSTAIN
17. GET IN TOUCH! Tobias Bohnhoff Appanion Labs GmbH ▪ Hopfenstrasse 11 ▪ 20359 Hamburg ▪ www.appanion.com Discover more at www.appanion.com Disclaimer: This whitepaper is based on research and data of the previously mentioned sources. All information presented were researched and prepared by Appanion with great care. For the presented data and information Appanion cannot assume any warranty of any kind. Information and data represent in selected cases individual opinions and may be in need of further interpretation as a basis for decisions. Advances in research may occur after the publishing of this document. Therefore, Appanion is not liable for any damage arising from the use of information and data provided in this whitepaper.