AUGUST 2021 The Spectrum of Artificial Intelligence Companion to the FPF AI Infographic Contents

The Spectrum of Artificial Intelligence Companion to the FPF AI Infographic

I. EXECUTIVE SUMMARY______4 A. Symbolic AI ______4 AUTHORED BY B. Machine Learning______5 C. Risks and Benefits of AI______5 Brenda Leong II. INTRODUCTION ______6 Senior Counsel & Director of Artificial Intelligence and Ethics III. FOUNDATION DISCIPLINES______7 Philosophy______7 Dr. Sara R. Jordan Ethics______7 Senior Researcher, Artificial Intelligence and Ethics Logic______8 Mathematics______8 Physics______8 IV. MODERN COMPONENTS______8 A. Data______8 B. Statistics______9 C. Design______9 Security______9 Hardware______9 COVER ART V. ARTIFICIAL INTELLIGENCE (NON-ML) – OVERVIEW______10 Artist: Dr. Lydia Kostopoulos A. Rules Based AI______10 B. Symbolic AI______11 Title: Luncheon of the Tech Enabled Boating Party i. Search______11 Year: 2020 ii. Planning and Scheduling______11 iii. Expert Systems______12 C. Computer Sensing______12 D. Robotics______13 E. Knowledge Engineering______13 F. Natural Language Processing______14 G. Risks and Benefits – Rules-based AI______14 VI. ARTIFICIAL INTELLIGENCE (MACHINE LEARNING)______15 A. Reinforcement Learning______15 B. Neural Networks______16 C. Deep Learning______17 D. Generative Adversarial Networks______17 E. Risks and Benefits – Machine Learning______17 VII. CONCLUSION______19

THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 1 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE Produced by

Artificial Intelligence (AI) is the computerized ability to perform tasks commonly associated with human intelligence, including reasoning, discovering patterns and meaning, FPF.ORG generalizing, applying knowledge across spheres of application, and learning from experience. The growth of AI-based systems in recent years has garnered much attention, particularly in the sphere of Machine Learning. A subset of AI, Machine Learning (ML) systems “learn” from the success or accuracy of their outputs, and can change their processing over time, with MACHINE LEARNING minimal human intervention. But there are non-ML types of AI that, alone or in combination, lie behind the real-world applications in common use. General AI — a human-level ML Algorithms improve through experience computational system — does not yet exist. But Narrow AI exists in many fields and applications where computerized systems greatly enhance human output or outperform humans at defined tasks. This chart explains the main types of AI, their relationships to each other, and provides specific examples of how they are currently appear in our day-to-day lives. It also demonstrates how AI exists within the timeline of human knowledge and development.

GENERATIVE ADVERSARIAL NETWORKS GAN Two NNs learn by fighting SYMBOLIC AI SA Human-readable logic problems AI USE CASES AND CONTEXTS DEEP LEARNING DL Multiple layers of FINANCE neural networks TAX COMPLIANCE A software platform that distills tax laws into a EXPERT SYSTEMS ES Complex solutions program, creates a personalized decision system, through reasoning REINFORCEMENT LEARNING and enables individuals to quickly and accurately RL Learning to complete a task file their taxes. Value of AI: Tax compliance requires complete accuracy. This efficient, interactive system that provides precise and logically connected results SEARCH S Steps from initial allows taxpayers to understand, confirm, and have state to goal confidence in the outcome. KE provides transparent and clear explanations. NEURAL NETWORKS ROBOTICS DEEP LEARNINGNN Learning by making R Multi-sensing and/or DL Multiple layers of DESIGN USER Types of AI: connections PLANNING & SCHEDULING mobile AI neural networks P&S Multi-dimensional strategies and action sequences EXPERIENCE KE NN NLP USER INTERFACE RULES BASED HEALTHCARE RB Deductions based on curated rules KNOWLEDGE ENGINEERING KE Rules based on human expertise AMBIENT CHARTING PHIL The use of background voice-to-text processing OSOPHY FOU during a patient/medical provider exchange to record those interactions into the patient’s chart, ND along with extracting tasks, symptoms, and ATION recommendations for further action as required. MA A Value of AI: Medical providers spend significant L TECHNO THEMA COMPUTER SENSING time documenting, with uneven outputs, as well as CS Human sense-based inputs difficulty in correlating between providers. Ambient ETHICS TICS systems encode conversations, target key phrases, LO NATURAL LANGUAGE PROCESSING GY NLP Understand, interpret, manipulate language and present a summary for provider edit/acceptance. STATISTICS PHYSICS DATA BUSINESS Types of AI: LOGIC ANAL ANALYTICS YSIS SECURITY SA DL NLP ENCRYPTION

MODELING TRACKING MOBILITY AND TRANSPORTATION SOCIAL MEDIA FORECASTING WORKPLACE MONITORING TURN-BY-TURN NAVIGATION SPEECH OR CONTENT MODERATION SUPPLY CHAIN MANAGEMENT Embedded systems can monitor physical and digital Location-based software that provides detailed Systems can facilitate human teams in identifying, Systems to improve traditional inventory and traffic, data usage, device management, and some instructions for travelers to reach a selected flagging, and deleting posts with defined, prohibited forecasting beyond historical/internal trend data, employee behaviors for efficiency and security designation, customizable mode of transportation, terms (such as “hate speech” or profanity). to weight and include external factors such as management of time, assets, and resources. multiple stops, services en route, and real-time Categorizing and selectively reacting based on weather, consumer sentiment, demographic trends, Value of AI: Monitoring enables necessary adjustments based on traffic, tolls, and weather. platform policies, usually embedded in analysis of portal traffic, stock fluctuations, and enforcement of data security policies and protocols. Value of AI: This is a “shortest path” problem human/computer systems for review and decision. service levels HARDWARE Also, systems can monitor and manage time solver, able to consider and weight variables such Value of AI: More efficient at scale than human-alone Value of AI: Systems can increase accuracy and reporting and project management tools, as well as as speed, cost, and personal preferences, and reviews. Additionally, well-designed systems can efficiency, as well as provide improved transparency ensuring appropriate supervision, training, and allow personalization based on repeated journeys, potentially adapt to variations in context, intent, and reliable, predictive analytics; enable aggregate support, including for remote workers as well as link to calendar and scheduling data, cultural norms, and user expectations more forecasting from individual impact up through and interactive prompts. consistently across platforms. regional levels. Types of AI: Types of AI: Types of AI: Types of AI:

RB CS NN S SA DL GAN KE NLP RL P&S ES KE ML I. Executive Summary

This paper is a companion piece to the FPF Spectrum several algorithmic designs that are generally consid- image) based upon the initial human programming. The of Artificial Intelligence (AI) Infographic (fpf.org/blog/ ered to be examples of Symbolic AI, including Search, B. Machine Learning other network, the “discriminator,” has been programmed the-spectrum-of-artificial-intelligence-an-infographic-tool/), Planning and Scheduling, and Expert Systems. to what the correct output should be (e.g., what the image to expand the information included in that educational The types of AI described above tell computers how to should look like). The discriminator evaluates the output, When computers are programmed to find a specific pat- resource, and describe how the graphic can be used sift through information according to rules and process- and critiques it. Initial outputs are likely to be extremely tern in a set of symbols and then to perform a designated as an aide in developing legislation or other regulatory es crafted by humans, such as language or mathematics. inaccurate. The discriminator’s feedback is then incorpo- action, we can say that the AI is engaged in a “search.” guidance impacting AI-based systems. We identify spe- Machine learning is different. Machine learning works rated, the generator continues to churn out results, and Planning and scheduling AI are what enable a comput- cific use cases for various AI technologies and show how because machines use an initial set of rules (program- the feedback loop continues until the generator produces er to take into account multiple dimensions that require the differing algorithmic architecture and data demands ming) to identify connections and patterns which they data that the discriminator believes meets the quality adjustments of strategies, such as collecting tokens in present varying risks and benefits. We discuss the spec- then use to internally edit their instructions or build ad- expectations. This GANs type of learning is what drives a video game (e.g. gaining points) while avoiding traps trum of algorithmic technology and demonstrate how ditional rules of their own. Computers using the results “deep fakes” and some entertainment uses of AI and will (e.g. losing lives). Search AI and planning and sched- design factors, data use, and model training processes of prior analyses to improve subsequent calculations or likely inform or improve other systems in the future. uling AI play important roles in the boring and hidden should be considered for specific regulatory approaches. minimize loss of performance is what makes machine parts of systems that provide many of the conveniences learning so powerful. Machine learning has improved Recent calls for regulation of AI-based systems exist in modern life. For example, supply chain management, traditionally difficult AI tasks, such as image recognition, within the complex landscape of contemporary AI pro- including airline cargo scheduling systems and “just and provides the ability to analyze constantly changing C. Risks and Benefits of AI grams within a variety of industries and applications from in time” restocking models rely on these programs. information flows for applications like social media con- agricultural crop growth detectors to automated book Expert systems identify solutions by combing through tent monitoring. The future of AI ultimately lies in the goals and systems recommendations.1 Some approaches focus on a need and combining multiple types and layers of information, towards which humans direct it. Responsible uses to cover “algorithms” generally while other initiatives using the reasoning and logic common to a particular Neural networks, the building blocks of some types of should include two primary foundations for AI: to further are narrower, suggesting regulation of AI specifically as profession or specialty, such as medicine or engineer- machine learning, learn by identifying patterns within advance human knowledge and to improve human lives. used in automated vehicles2 or medical devices3. This ing. An expert system can provide faster, more reliably input data to make new, internal, rules about the rela- AI is key to the future of knowledge in many scientific paper aims to assist the development of any of these accurate diagnoses based on personal health data, or tionships between the data and outputs. These systems disciplines and commercial technologies but carries approaches by making clear some of the terms, rela- design recommendations for pollution abatement given allow computers to process highly complex information accompanying risks that it will be applied unethically, or tionships, and functions of AI, and how they might be the environmental and industrial factors involved. quickly, sometimes approaching human levels of asso- designed unfairly, and that individuals and groups will impacted by differing restrictions. Any regulatory efforts ciation and “intuition.” These networks can be layered be worse off in specific or personal ways. However, the Building computers that can “see,” “listen,” “smell,” or should be made with the best understanding of how to process data through multiple programs sequentially potential benefits are powerfully significant, if sufficient “taste” as ways to evaluate their physical environment different systems, use cases, applications, and ultimately and repeatedly for more sophisticated analysis. When effort is applied to ensure fair and beneficial impacts for requires new approaches to computer sensing that individuals, will be affected. Despite the media focus on they are layered, they may comprise a “deep learning” a greater social good. generally rely on a combination including several other Machine Learning (ML), there are many other types of AI system. These are the systems trained to recognize ob- forms of AI as well. Computer sensing plays an import- AI systems operate across a broad spectrum of scale. preceding and operating alongside ML, all of which have jects in photos, evaluating for color values, edges, and ant role as a building block for creating augmented Processes using these technologies can be designed different attributes and aspects, and which will need commonly associated items so that the output value, intelligence used in advanced and assistive robotics. to seek solutions to macro level problems like environ- differently formulated regulatory controls or guidance. such as probability that a specific image is a canoe and mental challenges around undetected earthquakes, Teaching AI to reason using the same cognitive patterns not a cat, can be provided to the user. Many forms of artificial intelligence are the clear ana- pollution control, and other natural disaster responses as expert professionals involves teaching machines the logue to human thought processes or human physical Most recently, there are two newer forms of machine while they are also incorporated into personal level bases of professional common knowledge laid out as a actions which can be integrated within systems to help learning enabling powerful systems to achieve major systems for greater access to educational, economic, set of underlying rules for processing and establishing humans process or move faster, more precisely or advances: generative adversarial networks (GANs) and and professional opportunities. If regulation is to be expectations, knitting together the wealth of documents, consistently, and with more information than an individ- reinforcement learning (RL). Reinforcement learning is effective, it should focus on both technical details and rules, and common knowledge of professions. Known ual could. But, of course, AI systems can also process a key step in designing AI to independently learn hu- the underlying values and rights that must be protected as “knowledge engineering” systems, they use rules large amounts of data beyond what any human could man-like, goal-oriented, tasks. Reinforcement learning from adverse uses of AI, to ensure that AI is ultimately and pattern recognition to sift through data such as tax do, to identify patterns, make connections, and predict is what powered the system that learned to master the used to promote human dignity and welfare. codes, extract relevant patterns, and categories, and outcomes that are far beyond what conventional data game “Go.” The first ML system, which defeated the provide an expert’s guidance. For example, this analysis science and statistical methods could accomplish. human world Go champion 4 games to 1, operated on could formulate question and answer sets to guide an traditional machine learning processes. The next pro- individual through their tax return preparation, following gram was designed using reinforcement learning and A. Symbolic AI the appropriate steps for that individual’s finances. beat that original system, 100-0. Reinforcement learning Natural language processing (NLP) systems are some systems will likely power the next generation of robot- Traditional Rules-Based AI leading to Symbolic AI of the most common AI systems that people routinely ics, for purposes such as search and rescue missions (sometimes used synonymously) is the collection of all encounter. Powering home-based assistants, and various in complex environmental situations or high-capability methods in artificial intelligence research that are based devices or appliances; providing language translation home care assistants. on human-readable representations (“symbols”) of tools, predictive typing, autocorrect, question and answer GANs, the newest variation of machine learning AI mathematics, logic, and coded programming. Symbolic systems, and robotic speech; summarizing and analyzing being developed, are based on creating a pair of neural AI systems represent the first significant steps towards written texts, and comparing drafts for plagiarism; and networks that learn by attempting to better each other: designing machines to enable complex decisions or even writing independent, creative work, NLP combines first, the “generator” of the pair creates an output (e.g., an reason through complexity and uncertainty. There are ML with other forms of AI in everyday systems and tools.

4 FUTURE OF PRIVACY FORUM | AUGUST 2021 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 5 II. Introduction

This paper is a companion piece to the FPF Spectrum poorly managed or unforeseen risks caused damage to particular dish, a rules-based AI system is a set of of Artificial Intelligence (AI) Infographic, to further humans, the environment, or social good. commands such as if-then logic statements that Philosophy explain the information included in this educational tell a computer how to combine precise forms The second, narrower argument focuses on the un- resource, and in particular how it should be used of information to achieve a particular task or known risks that may arise due to the innate design Each of the foundational disciplines of artificial as an aide in developing any legislation around AI- outcome. Algorithms can be more or less exact aspects of AI that appear to be out of the control of intelligence discussed on the following page — based systems. in their outputs based upon the clarity of the the programmer or user of the system. Under this logic, mathematics, physics — were once part instructions we give. Just as a recipe written by If one of the many calls to regulate AI were success- argument, because some forms of AI are designed in of what we now call philosophy. In Western tra- a chef may call for a “pinch of salt” because most ful, what exactly would be regulated? In many ways, a way that humans cannot fully comprehend, explain, dition, philosophy is the love of wisdom; per the chefs have a common understanding of what a the answer to this question is “it depends.” Some or reproduce, we should use AI with regulatory-driven Greeks, to love wisdom is to seek the pursuit of knowledge. “pinch” is, an AI system that instructs a computer approaches start from the premise that AI is a discrete caution until we can be more confident of the reliability While knowledge can be both theoretical or practical, within to reason probabilistically or approximately may type of software technology, like an operating system, or accuracy of these systems. Both of these arguments the context of modern AI, the impulse to gain practical knowl- be used in cases where a “good enough” answer and assume regulations should focus on preventing un- assume there are unique aspects to AI-based systems edge to deploy to a particular goal generally dominates. helps augment human decision-making. Also fair impacts on things like hiring or credit scoring. Other that are not, or cannot, be addressed by existing like recipes, AI systems can range from simple Western philosophy is only one of the views that informs think- approaches define AI as an entire system of hardware legal protections, regulatory schemes, or traditional to extremely complex, with some of the most ad- ing about “intelligence” in computer systems. The Indian sub- plus software, like a robotic arm or a personal home as- policy approaches. vanced systems being combinations of multiple continent describes philosophy as a debate between humans, sistant, and should be regulated in ways based on safe- Our purpose is not to assert if or how AI systems should algorithms creatively combined. nature, and the gods. Wisdom through their texts is the ability to ty concerns, similar to modern connected automobiles. be regulated, but rather to provide a general under- deploy wit alongside theoretical or eternal truths on the way to As we will discuss, the contemporary spectrum of AI is standing of the variety of AI systems that may be behind gaining a full grasp of being, knowledge, and even nothingness. broad, and any call to regulate AI must align regulatory various applications and industries, and what the im- The competing traditions of Chinese philosophies reflect the controls that are appropriate to the context, and in light pacts might be, and to demonstrate that any regulatory III. Foundation salience of learning, character development, an appreciation of of the specific harms of the systems being considered. 4 action should be taken thoughtfully and in response to Disciplines uncertainty, and the importance of meaning as the product of Artificial intelligence is a term with a long history. the particular areas of concern. Blunt approaches that a lifetime of seeking. Other philosophical traditions also bear Meant to denote those systems which accomplish tasks seek to include any and all systems using AI-based AI and Machine Learning (ML) are relatively new upon how we measure the intelligence of AI systems today.5 otherwise understood to require human intelligence, AI models are considerably more likely to have undesired features of the scientific landscape, but they are is directly connected to the development of computer impacts, and/or unforeseen secondary consequences. built upon a long history of philosophical and science but is based on a myriad of academic fields and scientific developments, including areas such as Toward that end, this paper outlines the spectrum of disciplines, including philosophy, social science, physics, philosophy, ethics, logic, mathematics, and phys- Ethics AI technology, from rules-based and symbolic AI to mathematics, logic, statistics, and ethics. AI as it is de- ics. In more recent decades, scientific disciplines advanced, developing forms of neural networks, and that inform AI include data analytics, statistical signed and used today is made possible by the recent Ethics is a potentially controversial term with both seeks to put them in the context of other sciences and modeling, and cybersecurity and encryption. advent of unprecedentedly large datasets, increased positive and negative connotations. Historically, disciplines, as well as emphasize the importance of Furthermore, these systems cannot be evaluated computational power, advances in data science, machine ethics is a discipline that trains minds to consider security, user interface, and other design factors. Addi- without also including considerations about the learning, and statistical modeling. AI models include universal ideas about what it means to be human, tionally, we seek to make this understandable through hardware devices and networks, the underlying programming and system design based on a number of and addressing questions such as “what is good, how might providing specific use cases for the various types of AI logics and principles, and user interface and user sub-categories, such as robotics, expert systems, sched- I be good, or how might I do good?” But ethics can also and by showing how the different architecture and data experience designs. All of these areas contribute uling and planning systems, natural language process- generate concerns about people shifting boundaries around demands present specific risks and benefits. to the questions, answers, and analysis needed ing, neural networks, computer sensing, and machine good and evil, or subjectively defining right or wrong toward a to fully review present day AI. While some might learning. In many cases of consumer facing AI, multiple Across the spectrum, AI is a combination of various particular agenda or outcome.6 The vagueness about what is seem remote, or tangential, in fact the underlying forms of AI are used together to accomplish the overall types of reasoning. Rules-based or Symbolic AI is or isn’t included in ethics has also given rise to the use of syn- values and assumptions of the designers and performance goal specified for the system. In addition the form of algorithmic design wherein humans draft onyms and associated concepts for this area, such as “social users are key to understanding the contextual to considerations of algorithmic design, data flows, and a complete program of logical rules for a computer to responsibility,” “morals,” “values,” or “human rights” without implications of any particular AI system. programming languages, AI systems are most robust for follow. Newer AI advances, particularly in machine always being clear or consistent as to what each of these use in equitable and stable consumer uses when human learning systems based on neural networks, are able means. These various perspectives on ethical considerations designers also consider limitations of machine hardware, to power computers that carry out the programmer’s have flowed over into twenty-first century discussions around cybersecurity, and user-interface design. initial design but then adapt based on what the system the appropriate roles of industry, technology, and automation, can glean from patterns in the data. These systems can Two arguments are most commonly behind calls for all of which continue to influence contemporary debates score the accuracy of their results and then connect regulation. The first is that AI presents unique risks about ethics in AI. These now include robust arguments of those outcomes back into the code in order to improve to humans, the environment, or social and cultural ideas such as whether an AI system can “be” ethical (or uneth- the success of succeeding iterations of the program. values at a scale and scope beyond prior technological ical), whether it’s dependent on the applications and context, advancements. Under this argument, AI regulations A commonly used comparison for an algorithm is that it whether human labor and safety priorities should pre-empt should introduce mechanisms to identify and mitigate is like a recipe. This analogy applies well to rules-based technological efficiencies, and even ideas like whether robots potentially harmful outcomes, both tangible and intangi- and symbolic AI. Like a recipe instructs the cook to com- are independent entities with agency or rights.7 ble, as well as offer solutions to rectify situations where bine specific ingredients in a specific order to make a

6 FUTURE OF PRIVACY FORUM | AUGUST 2021 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 7 IV. Modern Components

A. Data Logic B. Statistics Security

“Data” is the organized record and presentation of A branch of applied mathematics often tailored into information. Modern “big data”12 in the digital world is Logic is a branch of mathematics based on specific fields (e.g., statistical mechanics, biostatistics), Security is an intrinsic element in the part of the trail created by people, businesses or other established rules and proofs. It is concerned statistics inform the way in which the inputs for AI sys- trustworthiness and reliability around AI organizational entities as they interact with sensors, with evaluating conditions and relationships tems are evaluated and how the outputs are interpret- systems, including hardware, software, and surveys, sites, and applications embedded in the array with statements such as “if, then.” Symbolic ed.17 Statistical modeling is the design of a way to make network components. AI can be both a vul- of internet or computer-based, or computer-connected, logic, modal logic, predicate logic, syllogistic logic, and sense of data through analysis. Without established nerability, and a part of the defense of computers sys- products, services, and features.13 Data in this context computational logic are all part of the study of how statistical theory, we would lack many of the essential tems against cyber intrusion. Data can be “poisoned” has measurable characteristics like volume, variability, humans record their ideas, reason about their innate concepts to explain how AI works or to evaluate how ac- at multiple stages in the life of a model, and models accuracy, and value, and also can have organizing prin- connections, and reach conclusions that can be rep- curately machines have identified patterns. Established themselves can be subverted. But AI-based systems ciples around protection, privacy, provenance, and pro- licated.8 In the areas of AI and machine learning, logic tools and measures such as variance, correlation, con- are also being integrated as powerful tools for defend- portionality.14 In discussions about AI and ML, personal plays an intrinsic role in the translation of human rules, fidence, probability, and hypothesis testing inform the ing against cyber attacks targeting datasets as well data is often categorized according to the sector from conclusions, or understanding into machine parameters algorithmic models in order to glean insights and inform as existing systems, such as energy grids, or defense which it is gathered, such as healthcare, financial, edu- and actions. Indeed, symbolic and rules-based AI could evidence-based reasoning.18 This reasoning drives how networks. Security concerns are a critical component cational, or consumer data. Data is also described based not exist in their known forms without logic.9 systems combine data to test representations of reality of all stages of system design and implementation. In on the technological associations, including sensor data, against one another, and provides the tools necessary addition to the traditional threats to data, networks, location data, or visualized data.15 Data can also be syn- to evaluate the results. and storage, AI models in particular are at risk from thetic, imputed, or meta (data about data). Data is not just inference, inversion, and extraction attacks — all of the invaluable input into many AI and ML systems, but Mathematics which potentially compromise the system, whether by also includes the information created by or describing C. Design revealing the data, or inappropriately influencing the the analysis inherent in those systems, and the outputs outcomes or future processing. and recommendations the systems produce. Outside of computer programming, AI and While not always the first area of consideration for AI machine learning can almost easily be The techniques necessary for processing data for uses systems, the physical design of AI systems includes defined as areas of applied mathematics.10 in AI and ML have spurred both theoretical and applied considerations of aesthetic value, creativity in physical Weaving together linear algebra, calculus, research in areas of data management, including the presentation, and the efficiency of human interface. combinatorics, graph theory, information theory, Hardware ethical processing of data (e.g., differential privacy) and Apple, for example, is particularly known for its focus probability, and vector analysis, machine learning in par- the technical processing techniques like dimensionality on beauty and intuitive function as part of its innate ticular depends on the mathematical manipulation and 20 From the physical building of “server reduction (e.g., principal component analysis).16 The design structure. “Design thinking” — a process for analysis of data. Mathematical thinking and the repre- farms,” to considerations of the environ- demand for AI-ready data has also spawned an industry brainstorming collaboratively — includes 5 stages: sentation of the world in equations and numerical terms mental impact of the their huge power for data engineering, a new form of AI-adjacent special- Empathize, Define, Ideate, Prototype and Test and are critical to explanations of AI and machine learning. requirements, AI systems occupy a ization that serves to expedite, but potentially compli- is deployed widely in the software and applications substantial place in the physical world, cate, explainability and transparency for new machine development landscape in user interface design and integral to providing the digital and virtual platforms we learning applications. user experience design. More than simple efficiency or functionality, however, design considerations for AI have all come to take for granted. Cleaning data, identifying bias in data, adding noise to systems must also include the understanding that they Modern AI, and in particular ML, are possible only data, deidentifying data, standardizing machine-readable Physics are integral components of applications that affect because of advances in computer processing speeds data, analyzing data, and “owning” data are just a few of the lives of individual people, communities, cultures, and associated hardware infrastructure, such as the many issues surrounding the collection, flow, and use and the environment. Whether focusing on personal Even without consideration of the emerg- high-density, environmentally controlled servers, of all of this information in modern digital systems. privacy, bias and fairness, civil rights impacts, or other ing, transformational impact on AI that may high-resolution monitors, and even peripherals like social impacts, design choices are a fundamental con- arise when quantum computing becomes HDMI cables. Progress in AI depends on continuous, sideration in any AI system. more widely understood and available, the reliable access to high performance computing hard- role of physics in artificial intelligence cannot be un- ware and transmission channels. Even for cloud appli- derstated.11 Problems associated with mining the data cations or virtual environments, there is a hardware in- from massive physics research projects (e.g., CERN, frastructure where applications reside, computations astronomy) have pressed machine learning experts to are performed, and system management carries out identify new methods for managing data at the scale the necessary planning in support of efficient distribu- of the universe, as well as that of the quant. At the end tion of AI processing.19 of the day, AI systems operate in the physical world and are bound and scoped by the physical restraints of human-scaled applications.

8 FUTURE OF PRIVACY FORUM | AUGUST 2021 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 9 V. Artificial Intelligence – Overview

Artificial intelligence covers many combinations of hard- Inference engines work by either “forward chaining” category includes. Search algorithms are used in many ware and software components, but at its core, means or “backward chaining” through the rules. In a forward systems that are designed to find the best path to a the set of actions that would normally be understood to chaining strategy, an inductive approach (moving from goal within a specific set of possible solutions. Search require human intelligence. The most dazzling examples individual facts to general theory) moves forward from algorithms help us to achieve goals such as reaching the of AI that appear in science fiction are identical or superi- data inputs to conclusory outputs by matching the data highest score on a game, finding the shortest or fastest or to generalized human cognitive and intuitive powers. to the best rule to identify a match. In a backward chain- route to a destination, or finding a recipe for vegan choc- Whether this type of AI, known as General (or Strong) ing strategy, a deductive logical approach (from general olate chip cookies. At scale, and for industrial purposes, AI is even possible remains speculative, with no clear theory to individual facts) moves backward from the Search is one of the tools to help optimize supply chains. idea of when such capability might develop. However, provided set of possible outputs, through the rules and Search algorithms are a useful category of AI, but they Narrow (or Weak) AI is defined as the operation by sys- data, to identify if a particular output can be supported treat problem-solving holistically and without account- tems performing particular functions in a specific con- as legitimate. Where a known set of possible outputs ing for constraints. text, application, use case, or circumstance. Thus, the or decisions is generally well characterized, such as in A search algorithm can be best pictured by thinking greatest chess player, and the world Go champion, are healthcare diagnostic systems, a backward chaining through a grid environment. In a grid, a search algorithm now AI systems. These systems, while far outstripping strategy is useful. Where there is a wider or uncertain A. Rules-Based AI takes your initial position (such as 0,0 or the origin on a human capabilities in one specific area, are fairly use- range of possible outputs from a system such as select- coordinate grid), applies the possible set of actions you less for almost any other purpose. While some aspects ing, combining, and providing the appropriate mix of Artificial intelligence has been part of the programming can take, such as rules for chess piece maneuvers, cal- of their programming and sub-routines might be reused medications, a forward chaining strategy is useful. landscape, built on the academic foundational disciplines culates what will happen after each action, weighs them to jump start other projects, it is not a system that can described above, since at least the 1940s. The earliest as possible solutions to achieving a goal state (such as intrinsically “learn” a new skill (function or application) forms of AI are still in use today, including within some of arriving at 10,0), and weighs the actions across the spec- without human involvement to edit and reapply that our most widely used systems. In fact, examples of rule- ified action cost function, such as each move costing 1 code, provide new data and training, and so on. based AI, such as those which make up the backbone point. Good search algorithms help an actor to find a Across the spectrum, AI is a combination of various types of navigation systems, are so commonplace we often do low-cost solution using the least amount of resources, of reasoning. Rules-based or Symbolic AI is that form of not think of them as artificial intelligence any longer. including the cost of calculating complex solutions.24 algorithmic design wherein humans draft a complete The initial drive to build artificial intelligence grew from For example, when a computer is taught to attempt early program of logical, connected commands for a computer the human desire to automate work that is repetitive, computer games, such as Pac-Man, it can be taught to to follow.21 Newer AI advances, particularly in machine tedious, dangerous, requires high levels of precision, achieve a high score by finding each dot on the way to learning systems based on neural networks, are able or is simply impossible for an individual or group of hu- the piece of fruit. But, searching to achieve a relatively to power computers that carry out the programmer’s mans given some combination of factors. Many of these simple goal like “collect cherries in Pac-Man” can be initial commands but then adapt their operations based tasks are dependent on heuristics — formal and informal made more difficult by increasing the complexity of the on what the system can glean from patterns in the data. rules — that “everyone” within that field or domain just rules of the game, such as the need to avoid dead-ends These systems evaluate their results and then connect “knows.” Some of the first examples of rules-based or and ghosts in the mazes. To teach a computer to fully those outcomes back into the code in order to improve a. heuristic AI emerged in fields like chemical analysis, play Pac-Man — where it not only collects fruits but also the success of succeeding iterations of the program.22 B. Symbolic AI infectious disease diagnostics, or oncology diagnostics. avoids dead-ends and ghosts — requires an additional What was relevant about these domains was the belief Recently, what the media and many discussions around The terms Symbolic AI and Rules-based AI are many form of symbolic AI, Planning and Scheduling. that the logical structure of the questions to be asked AI are most commonly referring to is actually machine times used interchangeably. Systems that automate and the specific types of information necessary to an- learning, as if those two terms were exactly interchange- processes in an imitation of human reasoning, such as swer those questions was both available and could be able. In fact, however machine learning as a specific through the use of logical statements or identification of ii. Planning and Scheduling symbolically represented to a computer. The rules for subset of AI. Explaining the specifics of machine learn- patterns, are a type of symbolic AI. Computers can be reaching “good” decisions, and the expert knowledge For more complex sets of variables that ing and the distinction between it and other types of AI designed to detect patterns such as sequences of let- of uncertainty or confidence in the utility of a specific include constraints, such as finding a route is one of the primary goals of this paper and infographic. ters, numbers or other symbols, and then to reproduce piece of information, could be systematically organized to a city airport without using highway, or All ML is a form of AI, but not all AI is ML. This is one of an arrangement of those symbols so that humans can and ranked to produce a reliable, consistent answer by a by a particular method such as train, an AI will normally the key takeaways to inform potential policymakers in evaluate them. There are various versions of this type of non-human program that was on par with, or superior to, be programmed by combining Search algorithms with their consideration of regulating AI systems — that they system, which we are using as the umbrella term for the that deduced or provided by the human expert. Planning and Scheduling programming. should correctly identify what type of systems are be- following three sub-categories: hind the functions and applications they are concerned Rules-based systems maneuver through data using Planning algorithms are advanced Search algorithms that with, and how those operate, so that standards, guid- 23 treat problem solving more like the way a human would. an “inference engine”; a set of logical commands i. Search ance, and restrictions can be targeted appropriately. according to which a computer interprets information These planning algorithms can account for a variety of and relates that information to the set of possible out- Search algorithms are used by many people situational constraints, such as incomplete information, puts (e.g., a probability score) in order to answer the in contemporary life. Since the beginning of time constraints, and non-redundancy of resources. questions asked of it. Because of the logic and sym- the age of internet search engines, “search” These types of models are used to design work plans, bolic structure of inference engines driving rule-based has come to be synonymous with online browser search strategic design, and logistics planning for tasks like systems, they can work both backwards and forwards. capabilities, but that is not the entirety of what this helping the Hubble Space Telescope arrange its scans of

10 FUTURE OF PRIVACY FORUM | AUGUST 2021 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 11 V. Artificial Intelligence – Overview (continued)

the universe.25 When this sort of program includes time components, they also include Scheduling algorithms to show which steps are dependent on other factors, and which may take place simultaneously versus those that must occur sequentially. Global supply chain manage- ment, metropolitan transportation planning, and energy distribution grids all use these forms of Symbolic AI.

iii. Expert Systems Some of the first Symbolic AI systems helped experienced practitioners in various professional domains to make, or teach others to make, complicated decisions, thus giving C. Computer Sensing D. Robotics E. Knowledge Engineering them the name “expert systems.” Expert systems identi- fy possible solutions by combing through and combining Perhaps the most intriguing use of AI and machine learn- Robotics is a field at the intersection of physics, engineer- Knowledge Engineering is a field of AI oriented to build multiple types and layers of information. Expert systems ing today is the ability to design and train computers ing and computer technology that produces machines systems that emulate the judgment and behavior of sift through information using the reasoning and logic to “sense” and then evaluate the physical environment that substitute for, augment, or replicate physical human a human expert by codifying knowledge as rules and common to a particular profession or specialty, such as around them. Computers can be designed with a range actions. They do not all, or even most, look humanoid. relationships between data. These systems represent medicine or engineering. of sensors that simulate seeing, hearing, smelling, even But in all their varied forms robots are gaining intellec- knowledge as directed acyclic graphs which are able tasting, and thus can be trained to collect, process, and tual and mechanical capabilities that are largely due to to express complex calculations and logical eligibility The key components of expert systems are an exten- respond to active or passive stimuli extending well be- the continued expansion of the AI systems powering rules. The graphs can be easily queried and and the sive, detailed “knowledge base,” the “inference engine,” yond that of human sensory perception in both scope them. They are built on Rules-based programming, and results reasoned to automatically produce a calculation and the available working memory. The knowledge and accuracy. These sensors might perceive and mea- likely incorporate Sensing in many variations, and in the or decision result. When reasoning using a Knowledge base is the network of rules, logical statements, and sure light (including infrared spectrums), touch, distance more complex systems, incorporate multiple Machine Engineering system, a backward chaining algorithm is domain specific knowledge and reasoning that defines (range), temperature, acceleration, and speed. Although Learning algorithms. typically used. Backward chaining starts from the goal the expertise of humans. The knowledge base is an early programmers assumed that having computers in- and works backward to determine what facts must be engineered product constructed to organize the rules, Like General AI, many of our ideas about robots come terpret their environments visually would be one of the asserted so that the goal can be achieved. norms, protocols, and standards of the specialty, in- from science fiction, and may mean different things to easiest tasks, creating the algorithms essential for com- cluding ranking and ordering them in sequence of the different people. However, robotics in an AI context is As an example, Knowledge Engineering powers the puter vision has proven to be an extraordinarily difficult information needed. This organization is done using a likely to be a much less comprehensive machine. The programs designed to provide individual and business challenge that now routinely combines aspects of rules- system of symbols comprehensible to both humans and working definition usually includes some sort of me- users with the ability to comply with the ever-changing based and symbolic AI together with machine learning. computers such as programming languages, dictionar- chanical device, with the specific abilities necessary to taxation rules and regulations. Because there are so ies for natural language processing, or established rules Furthermore, computers are being trained to “hear,” complete a physical task, in a particular environment, many and they change so frequently to greater and for maneuvering around a defined environment, such as “smell,” or even “feel.” Differentiating and classifying with specified parameters. Robots require a power lesser degrees, it would be nearly impossible for an in- a chess board. sounds is taught to computers in much the same way source of some kind and contain varying amounts of dividual to effectively identify, extract, and reconcile the that “seeing” is. A similar set of algorithms, combined programming — there are plenty of robots with no AI new rules against the old ones. To manage this at scale Thus, expert Systems augment the decision-making with neural networks (discussed under Machine Learn- involved, but more and more are incorporating at least for people and businesses generally, an ensemble of capacity of professionals in a specific topical domain by ing, below) are designed to analyze and classify sounds. some level of advanced model or system. Some are rule-based algorithmic approaches are adopted: Natural systematizing their most expert levels of knowledge into Teaching computers to “smell” means to classify the designed to operate independently, or in many cases, to Language Processing is used to review current laws and heuristics (rules) and designing systems that can apply molecules detected, and makes use of older forms of AI, carry out tasks while in contact with cloud-based com- extract pertinent information; graph algorithms, such as them to new situations or large sets of data inputs. These combinations of AI and machine learning, and more re- putational resources, or carry out specialty operations networked analysis, can show the relationship of new systems are also the basis for other forms of algorithmic cently, a newly designed form of neural networks called under a human’s direct control. rules to previous instances of the rule and also reflect reasoning, such as some forms of natural language the impact of other applicable rules. The information ex- “graph neural networks.” The scope of robotics applications is expanding quickly. processing, and can be combined with scheduling and tracted can be further processed into ontologies (repre- Twenty years ago, most robots were doing things like planning systems, for use cases like robotics, and even Beyond the physical senses, there is research designed sentations of abstract concepts) that establish the terms assembling cars in factories. These consisted mainly of in newly conceived situations where the solution to a to train computers to both perceive and demonstrate to encode for use in subsequent applications, such as mechanical arms tasked with routine, repeatable tasks problem or the best action to take is not immediately emotive behavior from or on par with humans. However, those which forecast taxable income streams and asso- for attaching or manipulating car parts. But by today, obvious even to the human observer. attempting to design models which can reliably identify, ciated revenue for states and the federal government. self-operating machines and self-propelled units explore mimic, or initiate emotive behavior remains in the early And thus the whole is created to guide a user through a extreme climates on Earth and other planets, assist war- stages, is currently imprecise and inconsistent, and even detailed but highly individualized process based on their fighters and law enforcement, and are increasingly used where accurate, triggers some of the most challenging own information, against the background of the most in many aspects of healthcare and hospitality services. ethical and social questions. current rules.

12 FUTURE OF PRIVACY FORUM | AUGUST 2021 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 13 V. Artificial Intelligence – Overview (continued) VI. AI – Machine Learning

navigation systems may be unable to optimize for the fastest Machine learning is distinguished from Rules-based AI in applications, and any associated risks.30 Nevertheless, route in a way that accounts for all user preferences, such that it is coded so that the machine learning system can transfer learning is increasingly common as companies as external safety considerations or for roads with a history adjust its manipulation of the data to improve on human seek to maximize their return on investment. Particularly of flooding. And like any AI system, adverse consequences designated metrics for accuracy and fit, as in the case of given the environmental costs associated with developing may arise from flaws in the model components or the data, making better recommendations. The systems identify new or state of the art machine learning algorithms, trans- whether from errors in the knowledge base or inaccuracies patterns in a dataset, edit inferential rules based on those fer learning is increasingly being explored.31 within the inference engine for knowledge engineering. patterns and connections, and then judge the fitness of Machine Learning is applied in areas like healthcare, re- that model’s outputs against the requirements set by the These systems’ highest value is most often as they tail, e-commerce, recommendation engines, self-driving human programmers. When we say that a machine “learns,” augment humans already carrying out these process- cars, online video streaming, IoT (connected devices, it means the program iterates the process enough times to es. Expert systems facilitate more accurate and faster voice assistants, connected home appliances), and trans- perform better and better, to make the model outputs more diagnoses and some treatment processes but have not portation and logistics, as well as many, many others. accurate against a previously set standard for success. replaced the need for human doctors. While domain knowledge can be carefully engineered, the process of Machine learning has a number of sub-categories as well. F. Natural Language Processing maintaining and expanding a knowledge base to account ML systems can identify patterns in data we already know for new information introduces complexity, overlap, and something about, such as when it classifies or predicts Designing computers to speak, as well as to understand the potential for errors within computerized expert sys- based on similarities of existing preferences, and can also speech and text, is one of the most fundamental functions tems. In the complex arena of medicine and diagnostics, derive conclusions from data we know nothing about yet, necessary for the general progression of AI systems, keeping a knowledge base current presents an ongoing such as when it clusters items based on unseen connec- and research on these various capabilities all involves challenge, with the potential risk that patients will not tions or even generates new data. While there are general some forms of natural language processing (NLP) — a receive the latest, best, or most effective treatments. methods for training machines and designing machine combination of Symbolic AI and Machine Learning. Be- learning systems, many programs are ultimately unique to There are also considerations like risks to physical safety cause language is itself a set of rules (grammar, syntax, context, designed for a specific problem. These programs when general planning algorithms are not appropriately verb forms, etc.), individualized language programs can predict, classify, categorize, mine, and learn from the data. tooled by human designers, such as those used to direct be organized to process data according to those spe- Even machine learning program that start out identically industrial robots or other automated industrial processes. cific rules to classify or predict language. For example, become unique because each has “learned” through its Likewise, planning algorithms such as those which might when a large body of text (corpus) is examined using a own data-driven experiences, influenced by the initial make up the logistics strategy for delivering pandemic A. Reinforcement Learning dictionary (a set of search terms of interest) or parts of weight and importance values assigned by the program- vaccines or other crisis response materials, may not pro- speech are tagged, the rules are used to construct the mers, and then modified over time with real data. outputs, whether that be restatements, translations, or a vide optimized plans if they are not written to sufficiently Some machine learning algorithms are designed with sentiment analysis score. NLP covers the full scope of account for all the challenges involved, such as reaching The primary types of machine learning training are super- a model of learning that is based on exploration and systems designed to understand text, perform analysis remote areas, prioritizing population cohorts for tiered vised and unsupervised learning, along with variations like trial-and-error without explicitly relying on existing data or translation, and interpret and create text, as well as receipt, and storage and expiration — some of which may semi-supervised, self-supervised, multi-task, and transfer inputs. These systems seek to reach a particular level of understanding spoken inputs, and speaking in return. be difficult to define in new contexts, or generate require- learning. Without delving into detail on each, the types accuracy by generating various outputs, checking them ments for which information is not available. As with all au- are essentially defined by what type(s) of data are used against expectations, and then continuously editing their tomated processing, the quality of the outputs will always to develop, train, or test the system. Supervised learning process to get closer and closer to a match between be driven by the quality of the design and the availability is a system using data that has been labeled by humans solutions discovered and goals. Various factors of the G. Risks and Benefits – Rules-based AI and accuracy of data inputs behind the model or system. (a process that can raise ethical issues all on its own).27 outputs — such as money, time, or other resources — are Unsupervised learning uses unlabeled data and exploits The newest designs of algorithmic systems are the given weighted values, and the algorithm is then tuned the connections that the computer identifies, largely Rules-based or Symbolic AI may seem simple or even various forms of Machine Learning (ML). As mentioned, to explore in such a way that gains it the maximum re- independently.28 Semi-supervised learning is in between old-fashioned compared to its “smart,” machine learning the attention and public focus on these systems as they ward by performing those actions that bring the highest these, when the system uses a small set of labeled ex- counterparts. However, many apps commonly used to- have begun to pervade everyday devices and systems reward values with the least errors. In other forms of amples and learns from those while also evaluating and day, as well as much of the AI used to power the contem- has been so prevalent in recent years, that for many reinforcement learning, the computer proceeds without analyzing the unlabeled parts of the same dataset. porary economy, make use of these variations of AI. The people, ML and AI are treated as one and the same. any prior knowledge of the environment, adding each power of these systems is in solving design challenges With the advent of mobile devices such as smart phones Just as humans can learn to perform new tasks simulta- new learned fact after each action and thus building its 32 for integrated circuits such as on computer chips, creating and tablets, machine learning has been implemented neously, machines can be taught how to “multi-task” or own model of choices that lead to rewards. spam filters for emails, and crafting workday schedules. 29 to provide everything from weather apps to video and to perform related tasks simultaneously. And transfer For example, a program designed to play a video game They can be applied to exceedingly complex problems. book recommendations, personalized healthcare and learning is when the functional skills of a machine learning may be designed where the reward is gold coins and the Most recently, for example, research into protein-folding fitness platforms, and programs to accelerate and sup- model are transferred from one domain, or use case/appli- errors are the loss of game lives. A reinforcement learn- structures, such as AlphaFold, build on the earlier work port employment and educational needs. The average cation, to another, with appropriate adjustments tailored ing system will figure out how to maximize the number of of assembly-line designs similar to the robotic assembly consumer may enjoy using these applications but will to the new topic and problem. This may be done for cost coins gathered while also reducing the number of lives of machine parts, are based on rules-based AI.26 likely struggle to describe what machine learning is, savings or other efficiencies, but can result in inaccurate or lost in a video game by repetitively exploring through However, as useful as symbolic systems are, they do have how it works, or where it contributes to the applications even ethically problematic outcomes if the updates made every possible set of choices, even without preset some limitations. For example, search and planning-driven found in their phones, televisions, or even appliances. for the new use were not sufficient to account for different guidance as to what its game behavior should include.

14 FUTURE OF PRIVACY FORUM | AUGUST 2021 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 15 VI. AI — Machine Learning (continued)

Reinforcement learning algorithms are what drive some systems can address business challenges such as sales of the notable forms of machine learning, such as when forecasting, consumer research, risk management pre- machines learn to play games, operate a flight simulator, dictions, and character recognition, among other things. or power a robot vacuum through a new environment.33 Neural Networks have multiple variations, including Con- And because reinforcement learning does not require volutional Neural Networks (CNN) and Recurrent Neural an existing large set of data on which to train or operate, Networks (RNN). Another form, Generative Adversarial it can avoid some of the technical and ethical challenges Networks (GANs) is discussed separately, below. involved with creating and maintaining such datasets. Convolutional Neural Networks (CNNs) are uniquely suited to operate on image data and are the algorithmic forms behind many facial recognition systems. CNNs use “kernel” filters in the input layer, meaning the image data is filtered in repeated and overlapping sections and then fed forward through the remainder of the system C. Deep Learning D. Generative Adversarial Networks as a highly refined “feature map” of the image. Most CNNs are feed-forward systems which learn to see the A subset of machine learning, Deep Learning simply re- Generative Adversarial Networks (GANs) are the newest features of an image by applying multiple filters in paral- fers to the number of total connections a program makes variation of ML being developed, and are based on a pair lel, something like viewing the same image in up to 512 in between the input layers and output layers, in that it of neural networks that learn by attempting to get the bet- ways for a complete mental model of the image. CNNs has more than three layers. Between input and output ter of each other. First, the “generator” of the pair creates “see” images somewhat like human artists see an object layers are “hidden” layers which perform additional an output (e.g., an image) based upon the initial human or a body — in ways that consider perspective, depth, calculations through refining the weights applied to the programming. The other network, the “discriminator,” has and lighting, and how these affect the representation. various features of the data. The output layer generates been programmed with criteria to evaluate the output B. Neural Networks For this reason, CNNs are used in objects detection and the final result to the user, usually a “recommendation,” (e.g., what the image should look like). The discriminator classification systems. CNNS can detect and differenti- along with the confidence level of the result (“yes, that considers the output, and critiques it — essentially de- Neural Networks (NN) are a type of machine learning ate discrete objects within an image, such as road signs, animal is a cat”). termining whether it is real, or correct. Early in the cycle, the initial outputs are likely to be far off from what is de- inspired by the human brain. They differ from other or cats. They commonly operate by finding the “edges” Deep learning models, particularly when applied to sired. The discriminator’s feedback is then incorporated forms of machine learning in that they are non-linear of different objects, and then comparing the arrange- CNNs or RNNs, comprise exceptionally powerful sys- 34 into the program and the generator continues to churn by design. A NN consists of a web of interconnect- ment of edges, with other images they’ve previously tems for complex tasks such as language translation that 37 out results, and the feedback loop continues, until the ed entities known as nodes; each node carries out a identified. (When this doesn’t work as accurately as could not be accomplished by rules-based AI, or by linear generator produces data that the discriminator believes simple computation, similarly to the neurons in the desired, undulating sand dunes may be mistaken for a machine learning where there is only one computational meets the quality expectations. In the case of an image, human brain. Neural Networks are a collection of the reclining human body, for example.) CNNs can compare pass of the data from input to output. The “looping” this is the generator “fooling” it into thinking a generated algorithms used in Machine Learning for data model- objects or faces in a 1:1 or 1:many fashion, depending on feature of deep learning models allow them to use both output is real. GANs are very new and their capabilities ling operating within this graph of nodes. Data passes the design of the system. In other words, if the system forward and backward computational paths through the are still being explored, but they are the systems used, through several layers of interconnected nodes, as is trying to find all the traffic lights in the image, then layers, making it possible to identify obscure or highly so far, to produce “deep fakes,” contemporary works of each node classifies the characteristics and information every item is simply “yes, traffic light,” or “no, not traffic nuanced patterns and relationships in the data.39 of the previous layer before passing the results on to light.” Other systems may want to identify as many items art, music, or writing in the style of long dead masters, or other nodes in subsequent layer. The layers of networks in the image as possible, without knowing in advance In addition to editing its own inferential reasoning to create entirely unique compositions by the AI system. pass the data through hierarchies of various concepts, what they might be. pathways, Deep Learning program systems can create which, like other machine learning models, allows them entirely new, additional hidden layers between the input In contrast, Recurrent Neural Networks (RNNs) learn to learn through evaluating their own errors. layer and output layer without human intervention. This E. Risks and Benefits — Machine Learning from data where timing and sequencing are important is part of what makes “transparency” or “explainability” The first, or input layer, is where data is received, such as features, to answer questions around forecasting — so challenging for the users, or even designers, of a Machine learning programs are only as good as their design data uploaded to a cloud service or live sensor data, and predicting changes in air quality, the next word in a deep learning program. The path through the computa- is then subject to a mathematical function that manip- sentence, or stock market risk management estimates and training. Just as humans risk injury when they rush to tional layers may be different even for consecutive data run great distances or perform complex sports maneuvers ulates the data for further use over multiple iterations. based on numerous highly variable factors. RNNs begin inputs, and the structure of the network processing is The input layer to a learning system is not simply data with data from the input layer passed through the hidden without proper preparation, machine models risk generating changing and evolving in real time. Following exactly faulty outputs, or even causing harm if they are not well-de- ingestion but is an active path of calculation. The input layers to the output layers, however, they include loops what happened (what computations were applied, in layer leads to (at least one) hidden layer, and finally to an within the hidden layers that mimic a type of short-term signed and if their training is not continuously monitored what sequence, and with what factors or features in- and updated. That these systems “learn” without additional output layer. Each layer contains one or more nodes. By memory of what has already been processed. RNNs can volved) can be difficult to observe or recreate. However, increasing the number or complexity of the hidden lay- also map a single input to multiple forms of output. This human programming does not mean they should operate there is ongoing research on explaining deep learning without human oversight. If a system initially learns from a ers, you increase the computational and problem-solving means that RNNs are well suited for problems such as systems, attempting to make progress toward seeing abilities (as well as the computational costs required).35 translation and voice classification where one input — a limited, badly organized, or insufficiently representative data into the “black box” to understand, describe, or replicate set — whether the limits are due to the size of the data, the Deep Learning (discussed below) is defined as NNs letter or a phoneme (smallest piece that distinguishes a what the systems did to reach a particular conclusion.40 with more than three layers.36 Neural Network-based word) — can lead to different outputs.38 variety of data, the completeness of the data, or the veracity of the data — it cannot work properly “in the wild.”

16 FUTURE OF PRIVACY FORUM | AUGUST 2021 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 17 VI. AI — Machine Learning (continued) VII. Conclusion

Machines learn under guidance, application, and eval- down, but the limitations of human bodies had not been AI is a field of science that encompasses technical, so- employment, and global efforts toward environmental uation by humans. If machines are applied to problems included in the code. Thus, if put into a real environment, cial, and policy considerations. As with any technology, health and sustainable practices are all made feasible or that do not truly match their design, then they will deliver such systems would cause severe harm to a pilot and costly there is no “neutral” system — the choices made of significantly more effective by AI. Healthcare in particu- poor results. If a machine is directed to find patterns in equipment. Likewise, reinforcement learning systems must what to automate; how to determine the type, style, and lar has been entirely disrupted and generally improved data that is too limited to fully represent the knowledge have the “common sense” programming to set general pa- features included; sourcing the data; and applying the by the capabilities of these systems. the computer needs to solve the problem presented, the rameters as well. Notably, some systems of security robots, system in specific contexts are all design choices that As individuals, private companies, governments and machine learning program will be underfit to the prob- which are hard-coded to avoid confrontational humans, carry implications for equity and harm. Automating hu- regulators seek to safely and responsibly implement lem.41 One example of the underfitting problem comes sought escape by diving into a nearby fountain.44 man functions and behaviors carries the inherent risks these programs into the everyday lives of individuals and from uses of machine learning for predicting complex of automating human errors and shortcomings. When CNNs are applied to image models, they can be infrastructures, they must understand the technology as it disease dynamics. If a machine is asked to identify the highly successful, reliably recognizing not only that a dog However, AI systems can also add specific and extend- functions within the particular industry or application, and major causal drivers for a disease, which nurses and is not a cat but that a Labrador retriever is not a Pekingese. ed value to many fields, providing efficiencies in cost carefully consider where and how appropriate restrictions therapists know is strongly influenced by social factors However, if poorly trained, they will distinguish wolves and time, as well as accuracy and reliability. Mobile ser- should apply. This paper will hopefully make that task but the machine is only given laboratory test data to from dogs simply on the basis of snow in the picture, or vices that provide banking, communications, and safer, easier, provide an understanding of the complexities, the learn from, then it will be underfit for the purpose of mis-categorize men and women of particular races, or with smoother travel, among others, would not be possible opportunities, and the limitations of these programs so helping medical teams manage the disease. Likewise, specific health conditions. When moving beyond image without it. Low cost goods, access to education and that they can be an overall benefit and boon to humanity. if a similar system is only trained on data from a single recognition to individual identification systems such as hospital system that uses a single set of vendors to facial recognition, models have been repeatedly shown provide health records and reports of diagnostic tests, to fail badly when the training datasets are insufficiently the learning system may be underfit when applied to diverse and representative of the faces that will be en- another hospital system using another vendor›s records. Endnotes countered once the system goes into production.45 There As has been covered extensively in the media and ac- are certainly facial recognition systems now that have ademia, machines that replicate historical human-run corrected this problem and are highly accurate across 1 Tanha Talaviya, Dhara Shah, Nivedita Patel, Hiteshri Yagnik, Manan 6 MacIntyre, Alasdair. A Short History of Ethics: a history of moral systems, and which are therefore trained on the historical all demographics without significant variance, but these Shah. 202. Implementation of artificial intelligence in agriculture philosophy from the Homeric age to the 20th century. Routledge, data from those systems, are assuredly going to perpetu- systems remain in the minority (and are the most expen- for optimisation of irrigation and application of pesticides and 2003; Blackburn, Simon. Being good: A short introduction to ethics. herbicides, Artificial Intelligence in Agriculture, Volume 4, Pages OUP Oxford, 2002. ate any biases or inequities of those systems. For example, sive) of those available for commercial use. Facial analysis 58-73. Available at: https://doi.org/10.1016/j.aiia.2020.04.002. 7 Schuelke-Leech, Beth-Anne, Sara R. Jordan, and Betsy Barry. predictive models such as those that are used by financial systems that attempt to categorize people by gender, or 2 National Highway Traffic Safety Administration. 2017. “Automated “Regulating Autonomy: An Assessment of Policy Language for organizations to predict loan risk, if trained solely on his- identify characteristics or emotions remain highly flawed Driving Systems: A vision for safety”. Available at: https:// Highly Automated Vehicles.” Review of Policy Research 36, no. 4 torical loan data, will exhibit the same historic biases tied for technical accuracy as well as ethically suspect. www.nhtsa.gov/sites/nhtsa.dot.gov/files/documents/13069a- (2019): 547-579 Jordan, Sara R., and Phillip W. Gray. “Clarifying the to race, educational institutions, or other discriminatory ads2.0_090617_v9a_tag.pdf concept of the “Social” in risk assessments for human subjects RNNs face the biggest challenges related to transparency, factors. That is, some customers will be privileged in their 3 Food and Drug Administration. 2021. “Artificial Intelligence and research.” Accountability in research 25, no. 1 (2018): 1-20. as discussed above, as they have an element of inscrutabil- receipt of loans while others are systematically excluded. Machine Learning (AI/ML) Software as a Medical Device Action 8 Priest, G. (2017). Logic: A very short introduction (Vol. 29). Oxford ity to their structure that makes them difficult to explain or Plan. Available at: https://www.fda.gov/medical-devices/software- University Press; Floridi, Luciano. Information: A very short In fact, in such systems, the biases can be perpetuated replicate. Explaining fully how the loops between layers in medical-device-samd/artificial-intelligence-and-machine-learning- introduction. OUP Oxford, 2010. even when the programming has attempted to minimize RNNs change the data and thus impact the output score re- software-medical-device 9 Apt, Krzysztof R. “Logic Programming.” Handbook of Theoretical such aspects, because the systems are capable of such mains an elusive goal for explainable AI. While it may not be 4 An important cultural note must be pointed out before examination Computer Science, Volume B: Formal Models and Semantics (B) nuanced associations (e.g. when zip codes become prox- of these foundational disciplines: Just as artificial intelligence 1990 (1990): 493-574. of immediate importance to know why an RNN offers one ies for socio-economic status or race). does not belong to one culture or computer programming to one translation of a phrase or another, it certainly is important to 10 Deisenroth, Marc Peter, A Also Faisal and Cheng Soon Ong. language, philosophy and mathematics do not exist only in cultures Mathematics for Machine Learning. Cambridge University Press. These bias challenges can sometimes be mitigated by ad- know that these neural network outputs suffer from fallibil- with a touch to the Northern hemisphere, Mediterranean or Atlantic (2020); Wilmott, Paul. Machine Learning: An Applied Mathematics dressing data quality, sufficiency and representativeness, but ities that mean they can make inaccurate or inappropriate Oceans, or Abrahamic religious traditions. Philosophy is a form Introduction. Panda Ohana. (2019). 42 46 of cultural expression common to cultures striving to answer the the biases are not always easy to identify and isolate. Alter- translations which humans should be cautious to trust. 11 Hogg, Tad, and Bernardo A. Huberman. “Artificial intelligence and questions: where did we come from, what are we, and how do we natives to improve these models include designing systems large scale computation: A physics perspective.” Physics Reports This delineation of the risks of various aspects of ML know? Mathematics is both the language of any market but also the using reinforcement learning (without the need for historical 156, no. 5 (1987): 227-310; Haugeland, John. Artificial intelligence: should not be read as outweighing the many benefits, language of abstract expression of natural constants. The brutal data), or synthetic data sets, but these are also not immune The very idea. MIT press, 1989. conveniences, cost savings, and efficiencies that have realities of history and material preservation mean that the set of to the inherent bias carried over by programmers and social knowledge recognized as philosophy and mathematics includes 12 Inside Big Data. 2020. “The 6 types of data everybody should know to already resulted from these systems, with more being avoid confusion”. Available at: https://insidebigdata.com/2020/12/25/ structures and must be evaluated and monitored carefully. disproportionately more of the thinking traditions written rather considered and developed all the time. From expanding than spoken or performed, written on paper rather than on stone, the-6-types-of-data-everybody-should-know-to-avoid-confusion/ While reinforcement learning models do not use historical access to credit and financial systems, improving health and composed in alphabetic and promiscuous languages that 13 Harford, Tim. “Big data: A big mistake?.” Significance 11, no. 5 (2014): data, they nevertheless sometimes produce undesirable care, supporting educational opportunities, and the entire translate words easily, even at risk of the loss of meaning. Due to 14-19; Neef, Dale. Digital exhaust: what everyone should know about big data, digitization and digitally driven innovation. Pearson outputs if their goals are not carefully designed and de- infrastructure of connected devices and home services, the geographical and historic situation of the authors of this report we rely on the written word, Western cannons of thinking, and the Education, 2014; Cunningham, McKay. “Next generation privacy: the fined. For example, during the trials for fighter pilot systems, ML has improved our lives in many ways. These gains will English language primarily. internet of things, data exhaust, and reforming regulation by risk of viewers watched reinforcement learning algorithms battle only continue to grow. Specifying the risks simply outlines 5 Jordan, Sara R., and Phillip W. Gray. The ethics of public harm.” Groningen Journal of International Law 2 (2014). one another by spinning in loops at forces impossible for the detailed and critical responsibility incumbent on those administration: The challenges of global governance. Baylor 14 Khan, M. Ali-ud-din, Muhammad Fahim Uddin, and Navaram Gupta. human pilots to withstand or by zooming towards the sky who develop, sell, and employ such systems to use (and University Press, 2011; Jordan, Sara R. “The innovation imperative: An 2014. “Seven V’s of Big Data: Understanding big data to extract or the ground to avoid one another.43 The system had been refrain from using) them responsibly, in alignment with analysis of the ethics of the imperative to innovate in public sector value”. Available at: http://www.asee.org/documents/zones/ zone1/2014/Professional/PDFs/113.pdf designed to maximize points based on avoiding being shot ethical values and a prioritization of human well-being. service delivery.” Public Management Review 16, no. 1 (2014): 67-89.

18 FUTURE OF PRIVACY FORUM | AUGUST 2021 THE SPECTRUM OF ARTIFICIAL INTELLIGENCE | COMPANION TO THE FPF AI INFOGRAPHIC 19 15 Halpern, Orit. Beautiful data: A history of vision and reason since 34 Willems, Heather. 2016. “The power of non-linear thinking”. 1945. Duke University Press, 2015. Available at: https://www.americanexpress.com/en-us/business/ 16 Mason, Richard O. “Four ethical issues of the information age.” MIS quarterly trends-and-insights/articles/power-non-linear-thinking/ (1986): 5-12; https://www.talend.com/resources/what-is-data-processing/ 35 B.K. Lavine, T.R. Blank, 2009. “3.18 - Feed-Forward Neural 17 Dangeti, Pratap. Statistics for machine learning. Packt Publishing Networks”, Editor(s): Steven D. Brown, Romá Tauler, Beata Walczak, Ltd, 2017. Comprehensive Chemometrics, Elsevier, Pages 571-586. 18 Bzdok, Danilo, Naomi Altman, and Martin Krzywinski. “Points of 36 Brownlee, Jason. 2020. “Crash course on multi-later perceptron significance: statistics versus machine learning.” (2018): 233. neural networks”. Available at: https://machinelearningmastery.com/ neural-networks-crash-course/ 19 Sciforce. 2019. “AI hardware and the battle for more computational power”. Available at https://medium.com/sciforce/ai-hardware-and- 37 Ziou, Djemel, and Salvatore Tabbone. 1998. “Edge detection the-battle-for-more-computational-power-3272045160a6 techniques-an overview.” Pattern Recognition and Image Analysis C/C of Raspoznavaniye Obrazov I Analiz Izobrazhenii 8: 537-559. 20 Apple, Inc. 2021. “Apple design resources”. Available at: https:// developer.apple.com/design/resources/ 38 For language processing tasks, attention models, such as transformers, are now used in the most common and best 21 An example of an algorithm for a specific domain represented as rules performing language systems such as “BERT” (Bidirectional Encoder for a human or computer to follow includes the Mayo Clinic urinalysis Representations from Transformers) or its permutations RoBERTa, adulterant survey algorithm, which is available at: https://www. DistilBERT, BioBERT, or BEHRT. While RNNs learn to identify mayocliniclabs.com/it-mmfiles/Adulterant_Survey_Algorithm.pdf patterns by observing sequences as dependent on time or sentence 22 Symbolic and neural systems are not walled off from one other. structure, meaning from start to finish, Transformers do not need Bader and Hitzler (2005) show how these two can be brought to process data sequentially to identify sequences as part of the together. Bader, Sebastian and Pascal Hitzler. 2005. “Dimensions output of a transformer model. This means that transformers are of neural-symbolic integration—A structured survey”. Available at: able to process data in parallel and without some of the problems of https://arxiv.org/pdf/cs/0511042.pdf RNNs. In layman’s terms, transformers can interpret the whole book 23 PCMag. 2021. “Inference engine”. Available at: https://www.pcmag. by reading all of it at once, rather than page by page or chapter by com/encyclopedia/term/inference-engine chapter, relying on a mental model of time or space development 24 Completeness is when an algorithm will find a solution when there such as how humans learn a story by watching the chapters unfold. is one or will report failure when a solution cannot be identified. Transformers can absorb all of the information without such a mental ACKNOWLEDGMENTS Optimality is identification of the lowest cost solution among all model. See for more: Liu, Yinhan, Myle Ott, Naman Goyal, Jingfei Du, options. Complexity is the amount of time or memory space needed Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Many thanks to the supporters of Future of Privacy Forum, especially members of to perform a search. and Veselin Stoyanov. 2019. “Roberta: A robustly optimized bert pretraining approach.” Available at: https://arxiv.org/abs/1907.11692; the Artificial Intelligence and Machine Learning Working Group, including industry 25 Samson, Roberto J. 1998. “Greedy search algorithm used in the or Staliūnaitė, Ieva, and Ignacio Iacobacci. 2020. “Compositional and automated scheduling of Hubble Space Telescope activities”. representatives, academic experts, and civil society advocates, who all contributed to Lexical Semantics in RoBERTa, BERT and DistilBERT: A Case Study on Proceedings of SPIE 3340, Observatory Operations to Optimize this paper by way of suggestions, feedback, and technical details. CoQA.” Available at: https://arxiv.org/abs/2009.08257 Scientific Return10.1117.12.316498. 39 A very simple analogy to describe the difference between an input 26 Callaway, Ewan. 2020. “’It will change everything’: DeepMind’s For further information about Future of Privacy Forum and its work on public policy layer and loading data is the difference between watching a 3D film AI makes gigantic leap in solving protein structures”. Available at: regarding Artificial Intelligence and Machine Learning, please contact [email protected]. with the glasses and without. The dazzling 3D features appear with the https://www.nature.com/articles/d41586-020-03348-4 red-blue filter of the glasses but do not when the glasses are not worn. 27 Savage, Saiph. 2020. “AI needs to face up to its invisible-worker 40 Montavon G., Binder A., Lapuschkin S., Samek W., Müller KR. (2019) problem”. MIT Technology Review. Available at: https://www. Layer-Wise Relevance Propagation: An Overview. In: Samek W., technologyreview.com/2020/12/11/1014081/ai-machine-learning- Montavon G., Vedaldi A., Hansen L., Müller KR. (eds) Explainable crowd-gig-worker-problem-amazon-mechanical-turk/ AI: Interpreting, Explaining and Visualizing Deep Learning. Lecture 28 Raw word clouds, which are representations of term frequency (word Notes in Computer Science, vol 11700. Springer, Cham. https://doi. count) and where no terms have been removed or no color coding org/10.1007/978-3-030-28954-6_10 applied, are a visual representation of unsupervised learning. These 41 Jabbar, Haider K. and Rafiqul Z. Khan. 2015. “Methods to avoid are not always particularly helpful and the frequency of seeing over-fitting and under-fitting in supervised machine learning”. word clouds seems to be on the wane. M. Hearst, E. Pedersen, L. P. Available at: https://www.researchgate.net/profile/Haider_Allamy/ Patil, E. Lee, P. Laskowski and S. Franconeri, 2019. “An Evaluation of publication/295198699_METHODS_TO_AVOID_OVER-FITTING_ Semantically Grouped Word Cloud Designs,” in IEEE Transactions AND_UNDER-FITTING_IN_SUPERVISED_MACHINE_LEARNING_ on Visualization and Computer Graphics. http://dx.doi.org/10.1109/ COMPARATIVE_STUDY/links/56c8253f08aee3cee53a3707.pdf TVCG.2019.2904683. 42 Lardinois, Frederic. 2017. “Intuit launches QuickBooks Capital, a 29 Xu, Yichong, Xiaodong Liu, Yelong Shen, Jingjing Liu, and Jianfeng small business lending service powered by AI”. Available at: https:// Gao. “Multi-task learning with sample re-weighting for machine techcrunch.com/2017/11/07/intuit-launches-quickbooks-capital-a- reading comprehension.” arXiv preprint arXiv:1809.06963 (2018). small-business-lending-service-powered-by-ai/ 30 Jamshidi, Pooyan & Velez, Miguel & Kästner, Christian & Siegmund, 43 Defense Advanced Research Projects Agency. 2020. Norbert & Kawthekar, Prasad. 2017. Transfer Learning for Improving Model “AlphaDogFight Go Virtual for Final Event”. Available at: https:// Predictions in Highly Configurable Software. 10.1109/SEAMS.2017.11. www.darpa.mil/news-events/2020-08-07 31 Sharir, Or, Barak Peleg, and Yoav Shoham. 2020, “The Cost 44 Garun, Natt. 2017. “DC security robot quits job by drowning of Training NLP Models: A Concise Overview.” arXiv preprint itself in a fountain”. Available at: https://www.theverge.com/ arXiv:2004.08900. tldr/2017/7/17/15986042/dc-security-robot-k5-falls-into-water 32 Bhatt, Schweta. 2018. “5 things you need to know about 45 Cornejo, Jadisha Yarif Ramírez et al. “Down syndrome detection based reinforcement learning”. Available at:. https://www.kdnuggets. on facial features using a geometric descriptor.” Journal of medical com/2018/03/5-things-reinforcement-learning.html imaging (Bellingham, Wash.) vol. 4,4 (2017): 044008. doi:10.1117/1. 33 Brandon, John. 2016. “Why the iRobot Roomba 980 is the greatest JMI.4.4.044008; lesson on the state of AI”. Available at: https://venturebeat. 46 FPF MasterClass Webinar, “Machine Learning and Speech,” com/2016/11/03/why-the-irobot-roomba-980-is-a-great-lesson-on-the- December 2020, presentation by Professor Marine. Carpuat, https:// state-of-ai/ www.youtube.com/watch?v=uRJsTy50Sqs

20 FUTURE OF PRIVACY FORUM | AUGUST 2021 The Future of Privacy Forum (FPF) is a catalyst for privacy leadership and scholarship, advancing responsible data practices in support of emerging technologies. FPF is based in Washington, DC, and includes an advisory board comprising leading figures from industry, academia, law, and advocacy groups.

1400 EYE STREET NW | SUITE 450 | WASHINGTON, DC 20005 [email protected] | 202-768-8950