Cognitive Systems For Every Occupation

Jim Spohrer DRAFT  05/25/2014 07:13 AM

Cognitive Systems for Every Profession

The Dictionary of Occupational Titles (DOT) contains hundreds of short paragraph description of occupations.  For example, Architect (001.061-010):

Researches, plans, designs, and administers building projects for clients, applying knowledge of design, construction procedures, zoning and building codes, and building materials: Consults with client to determine functional and spatial requirements of new structure or renovation, and prepares information regarding design, specifications, materials, color, equipment, estimated costs, and construction time. Plans layout of project and integrates engineering elements into unified design for client review and approval. Prepares scale drawings and contract documents for building contractors. Represents client in obtaining bids and awarding construction contracts. Administers construction contracts and conducts periodic on-site observation of work during construction to monitor compliance with plans. May prepare operating and maintenance manuals, studies, and reports. May use computer-assisted design software and equipment to prepare project designs and plans. May direct activities of workers engaged in preparing drawings and specification documents.

A system of professions exist in business and society (Abbot 1988).  Professions have jurisdiction over types of problems and their solutions, and reflect the division of expert labor.   All professions compile a body of knowledge and practices that define them, and university faculty extend and teach that codified knowledge.

We are at the dawn of an era where every professional will have one or more associated cognitive systems (see Appendix I below).  Cognitive systems ingest massive amounts of data, learn, permit natural language interactions, and provide levels of confidence in their recommendations.

In the era of smart system of systems, cognitive computing and other advances at long last make it possible to have cognitive systems capable of being true cognitive assistants.  Anyone lucky enough to have (or had) a great executive assistant knows the amazing boost for measures like productivity, quality, compliance, and innovativeness.

Productivity: From travel to organizing meeting and emails, to finding that presentation or contact from last year or three years ago.

Quality: Organizing and integrating feedback on draft documents, presentations, to following-up with thank-you’s and closing the loop.

Compliance: Any organization runs on a dozens of details, annual reports, and double checking compliance with process, procedures, policy and even regulations..

Innovativeness: By helping on all the above items, there is more time to think and interact with others on new topics of potential future value.

Up to this point in history, most people have never had the benefit of a great executive assistant, so they have to do-it-all-themselves, adopting good organization skills and disciplines to stay on top of everything needed in their professional and private lives.   A very few people have been lucky enough to afford both executive assistants at work and in professional/public life, and personal concierges in their home/private lives.   Because most people without personal assistants have both public and private roles, instead of dedicated assistants for each role, associated with each societal role, they have colleagues, friends and/or family who lend a hand to help out when needed.  Trusted relationships provide needed assistance for most people.

The best assistants also help those they serve “to up their games” and adopt better organization skills and disciplines over time.   Assistants who allow those they serve to remain scattered and undisciplined can “cover it up” for a shot period, but ultimately without elevating the efficiency and effectiveness of those they serve, growth, development, innovation, and advanacement suffer.  Like good coaches, good assistants understand that in the long run creating debilitating dependency is not what excellent service is all about, but instead growing capabilities on both sides of the service relationship is what leads to sustainable relationships and rewarding lives.  Anything less is merely transactional, not relational.  The exception, of course, is end of life assistance, where capabilities of the one being served dwindle over time, and the assistant must do more and more right up to the end.

In this short essay, the focus is “ProfessionCogs” designed to augment the intelligence of doctors, lawyers, engineers, journalist, and other professionals, however first let’s understand “Cogs” a little better.

Cogs: Learning, Language, and Levels

“Cogs” are human-made cognitive systems capable of being true cognitive assistants.  Cognitive assistants have three types of basic capability that we can summarize as learning, language, and levels.   We expect our cognitive assistants to not only remember the past, but learn from experience as well.  Furthermore, we expect to interact with cognitive assistants in natural ways, in written language at a minimum and ideally in spoken language with ability to use prosody, gestures, and even facial expressions to convey meaning naturally in multi-person interaction contexts, such as meeting and conference calls.  Finally, cognitive assistants must help us weigh alternatives and understand the level of confidence associated with different possibilities and recommendations.

So learning, language, and levels get cognitive systems to the point where they can be really useful.  For example, IBM’s Jeapordy! winning Watson system demonstrated all three capabilities.  Learning from previous correct and incorrect responses (training data), using a broad range of natural language terms and phrases (Wikipedia-type breadth), and levels of confidence in its responses (shown to the audience during the competition).  It is interesting to note that while no single IBMer could defeat the two most winning Jeopardy! players, the Watson Jeopardy! system, a very specialized “Cog,” which was developed by IBM and several university partners, could and did win.  “Cogs” can do some things that their developers or owners cannot do.

Cognitive Systems: A Diverse Set of Entities

In general, the set of all cognitive systems is a very broad set of entities. The term itself “cognitive systems” is somewhat hard to define clearly, simply, and precisely.   The set of all cognitive system entities that one could study include both biologically evolved entities as well as human-made, or artificially intelligent entities.   Cognitive science, the interdisciplinary field that studies the mind and it processes, includes researchers with expertise in both the evolution and function of brains as well as the design and practical applications of increasingly sophisticated artificial intelligence componentry.   Both fields provide methods for studying the physical structures that give rise to diverse cognitive processes (functions and behaviors).   Cognitive science and artificial intelligence are both also concerned with collective capabilities of interconnected networks of simpler cognitive systems that form larger structures – from social insects to cities to multi-agent systems of flying smart robots.

ProfessionCogs for Diverse Professions

In large organizations, ranking the performance of individuals and teams is a common practice.  From the best performers in a role to the weakest performers in a role, quantitative and qualitative assessments are used to evaluate performance.   Those with the lowest ranking have the greatest potential for improved performance, and therefore boosting their performance or replacing them with more qualified employees can be key to improving overall organizational performance.

Making Progress

Which brings us to the question of what is the best way to make progress going forward?

Definitions of cognitive systems make reference to human-level capabilities, such as the abilities to sense and respond to the world in intelligent ways.   However, cognitive science and artificial intelligence researchers include ants and thermostats as examples of cognitive systems, because they sense and respond to their environment.

Human-made systems with capabilities to sense and react to their environment have been around for years.

In sum, ProfessionCogs are cognitive systems/cognitive assistants designed to help professionals do their jobs better.

Further readings

There is a lot out there.  And more all the time.

Abbott, A. (1988). The system of professions: An essay on the division of expert labor. University of Chicago Press.  URL: http://psycnet.apa.org/psycinfo/1988-97883-000

====

Appendix 1: By Stephen Hamm, Cognitive Computing Overview

Cognitive Computing for Smarties, by Stephen Hamm

What is cognitive computing?

Cognitive computing enables the next level of partnership between people and computers to augment human intelligence, boosting the productivity and creativity of individuals and teams, thereby transforming industries and professions.

These systems ingest vast amounts of data, learn from their interactions with people and information sources, reason about their their level of confidence in derived knowledge, and interact using language and other means that are more natural to us.

Cognitive computing techniques provide the building blocks for iteratively developing increasingly sophisticated systems which help us to make better, faster decisions in our personal and professional lives.

The science behind cognitive computing:

To demonstrate cognitive computing in action, IBM built the now famous Jeopardy! TV game show winning machine known as Watson.

This history-making machine is capable of searching encyclopedic collections of information for potential answers to questions, ranking answers based on its confidence level in them, and pressing a button if it has enough confidence in its top-rated answer—all in less than 3 seconds.

IBM scientists worked for years to combine and innovate techniques from a number of computer science-related fields including machine learning, data mining, natural language processing, knowledge representation, text-to-speech synthesis, operations research, decision-making, game theory, cognitive science, psychology, linguistics, and more.

In universities, scholars typically pursue these fields in relative isolation. The Watson breakthrough came, in part, because IBM scientists and engineers combined the disciplines in new ways with a grand challenge goal always in mind.   Nevertheless, the breakthrough would not have been possible without the assistance of university researchers, and open frameworks such as Apache UIMA.

The annals of artificial intelligence, now include three game-playing machines from IBM: Watson Jeopardy! (2011), Deep Blue (Chess, 1997), Samuel’s Checker Player (1956).

However, the best is yet to come.

Cognitive systems will be able to…

–Understand multiple languages.
— Reason about levels of confidence in their derived knowledge.
–Converse with people in spoken dialogues.
– Multimodal interactions (see, gaze, facial expressions, hear, small, taste, touch, feel, empathy)
–Understand how professionals think—such as doctors and lawyers.
— Understand facial expression, voice, sensory information, and build deeper user models
–Help people make better decisions, learn complex material faster, make discoveries and create new knowledge.

And, they will get smarter over time.

Of all these capabilities, learning is key.

Like people, cognitive systems exhibit three types of learning over time.

Optimization: Learning to use existing knowledge more efficiently for specific tasks.

Education: Learning from other knowledge sources, people, books, the web, and other cognitive systems.

Discovery: Learning new surprising derived knowledge.

For now, most of the algorithms (cognitive computing techniques) for optimization, education, and discovery are provided manually by research scientists and engineers programming machines.   However, as the amount of knowledge in machine-readable form grows, knowledge itself with become a new form of big data for cognitive systems to use to derive new knowledge and algorithms in partnership with people and organizations that can benefit from this new knowledge.

As cognitive computers for all professions get smarter over time they will help people improve their performance as well.   With cognitive systems everyone can eventually have a combined expert tutor and cognitive assistant.   As professionals exhibit new best practices, their cognitive assistants will notice and learn, ultimately contributing to the body of knowledge for every profession.

In sum, cognitive computing enables the next level of partnership between people and computers to augment human intelligence, boosting productivity and creativity of individuals and teams, thereby transforming industries and professions.

Appendix II: Building Useful “Cogs” Is Hard Work, But No Longer Impossibly Hard

Why now?  Over the years, AI (Artificial Intelligence) researchers have come to appreciate just how hard the problems they have been working on actually are.  Overly optimistic claims in the past have contributed to so called “AI Winters” where funding dried up formost  AI proposals.  So rightly so, many skeptics are asking “Why Now?” to the claims that an AI Renaissance is underway.   Building these systems is still very hard work, but no longer impossibly hard.

Building “Cogs” (cognitive systems/cognitive assistants) that can answer questions is hard work.  However, for the first time in history, Linked Data on the web makes it feasible for many tasks.  As the amount of Linked Data on the web increases, some of the traditional very hard natural language processing tasks that early AI (artificial intelligence) systems attempted can be addressed.  The WWW’s Linked Data provides a practical solution to several early AI problems, such as (1) comprehensive online data sets, (2) knowledge representation at scale, (3) combinatorial explosion of inferences, and (4) insufficient memory or compute resources.

Jeopardy! is an interesting test of natural language and breadth of knowledge.  The correct response is always an entity (in the form of a question “What is X?”) that fits within a category (e.g., “Famous People”) and is referred by a clue.  In the age of Wikipedia, Wiktionary, Wordnet,Web Pages, and Linked Data, the words and phrases in the typical clue link to documents which contain the correct response in most cases within one hop, or in some cases a relatively small number of hops – three to five.  The diagram below illustrates what is meant by a “hop” from one web-page to the next (see Wiki Linked_Data for a better understanding).   Linked Data is a practical solution to the “combinatorial explosion” problem faced by early AI (Artificial Intelligence) systems.  Sometimes the correct response may require finding a set and performing some linguistic or numeric “Calculation” (e.g., Jeopardy Category: Letters of the Alphabet, Clue: Most Common State Name First Letter; Correct Response “M”).  This procedural knowledge common in word puzzles is part of what make Jeopardy! challenging for contestants.    Most Jeopardy! correct response includes an entity for which there is a web page, and a small number may include an entity for which there is no web page, such as result of a linguistic or numeric “Calculation.”

In multiple-choice grade level reading comprehension tasks, the correct response is most often an entity that is directly referred to in the answer passage, or in a web page document linked to a word or phrase in the answer passage by one hop, or in some cases a relatively small number of hops – three to five.  Again, Linked Data is a practical solution to the “combinatorial explosion” problem.

Professional certification can be thought of as grade level reading comprehension for a very long answer passage (one or two dozen books).

Professionals or experts are also often called on to present arguments for and against multiple future options.   Unlike tasks that require finding a single entity answer, there is no single best or right answer, only a list of possibilities with pro/con statements of support rank ordered for relevance to each option.  For example, expert debaters perform these types of tasks when they research issues and build cases.  Lawyers also perform these types of tasks.

Also, unlike tasks that require finding a single entity response, some tasks require summarization of events, meetings, publications, bodies of research.  For example, journalists are often confronted with these tasks, or textbook writers creating material that is grade-level appropriate.

Problems in pattern recognition and robotics are also still very hard, but many specific tasks are no longer impossibly hard.

Appendix III:  Non-Technical Issues Are Also Hard Work, But Not Impossibly Hard

As the range of tasks that cognitive systems can address there are three types of non-technical concerns: (1) threats to privacy, safety and security (??? by governments, businesses, criminals, terrorist, etc.), (2) threat to job security (Byrnjolffson & MacAfie,), (3) threat to species security (Weizenbaum, Joy, etc.).   Based on the law of comparative advantage, getting all entities to “up their game” creates the most individual and collective benefits.

Leave a Reply