Feeds:
Posts
Comments

Posts Tagged ‘Learning’

Human thinking processes are rich and variable, whether in search, problem solving, learning, perceiving and recognizing stimuli, or decision-making. But people are subject to limitations on the complexity of their computations and especially the capacity of their ‘working’ (short-term) memory. As consumers, they frequently need to struggle with large amounts of information on numerous brands, products or services with varying characteristics, available from a variety of retailers and e-tailers, stretching the consumers’ cognitive abilities and patience. Wait no longer, a new class of increasingly intelligent decision aids is being put forward to consumers by the evolving field of Cognitive Computing. Computer-based ‘smart agents’ will get smarter, yet most importantly, they would be more human-like in their thinking.

Cognitive computing is set to upgrade human decision-making, consumers’ in particular. Following IBM, a leader in this field, cognitive computing is built on methods of Artificial Intelligence (AI) yet intends to take this field a leap forward by making it “feel” less artificial and more similar to human cognition. That is, a human-computer interaction will feel more natural and fluent if the thinking processes of the computer resemble more closely those of its human users (e.g., manager, service representative, consumer). Dr. John E. Kelly, SVP at IBM Research, provides the following definition in his white paper introducing the topic (“Computer, Cognition, and the Future of Knowing”): “Cognitive computing refers to systems that learn at scale, reason with purpose and interact with humans. Rather than been explicitly programmed, they learn and reason from interactions with us and from their experiences with their environment.” The paper seeks to rebuke claims of any intention behind cognitive computing to replace human thinking and decisions. The motivation, as suggested by Kelly, is to augment human ability to understand and act upon the complex systems of our society.

Understanding natural language has been for a long time a human cognitive competence that computers could not imitate. However, comprehension of natural language, in text or speech, is now considered one of the important abilities of cognitive computing systems. Another important ability concerns the recognition of visual images and objects embedded in them (e.g., face recognition receives particular attention). Furthermore, cognitive computing systems are able to process and analyse unstructured data which constitutes 80% of the world’s data, according to IBM. They can extract contextual meaning so as to make sense of the unstructured data (verbal and visual). This is a marked difference between the new computers’ cognitive systems and traditional information systems.

  • The Cognitive Computing Forum, which organises conferences in this area, lists a dozen characteristics integral to those systems. In addition to (a) natural language processing; and (b) vision-based sensing and image recognition, they are likely to include machine learning, neural networks, algorithms that learn and adapt, semantic understanding, reasoning and decision automation, sophisticated pattern recognition, and more (note that there is an overlap between some of the methodologies on this list). They also need to exhibit common sense.

The power of cognitive computing is derived from its combination between cognitive processes attributed to the human brain (e.g., learning, reasoning) and the enhanced computation (complexity, speed) and memory capabilities of advanced computer technologies. In terms of intelligence, it is acknowledged that cognitive processes of the human brain are superior to computers inasmuch as could be achieved through conventional programming. Yet, the actual performance of human cognition (‘rationality’) is bounded by memory and computation limitations. Hence, we can employ cognitive computing systems that are capable of handling much larger amounts of information than humans can, while using cognitive (‘neural’) processes similar to humans’. Kelly posits in IBM’s paper: “The true potential of the Cognitive Era will be realized by combining the data analytics and statistical reasoning of machines with uniquely human qualities, such as self-directed goals, common sense and ethical values.”  It is not sufficiently understood yet how cognitive processes physically occur in the human central nervous system. But, it is argued, there is growing knowledge and understanding of their operation or neural function to be sufficient for emulating at least some of them by computers. (This argument refers to the concept of different levels of analysis that may and should prevail simultaneously.)

The distinguished scholar Herbert A. Simon studied thinking processes from the perspective of information processing theory, which he championed. In the research he and his colleagues conducted, he traced and described in a formalised manner strategies and rules that people utilise to perform different cognitive tasks, especially solving problems (e.g., his comprehensive work with Allen Newell on Human Problem Solving, 1972). In his theory, any strategy or rule specified — from more elaborate optimizing algorithms to short-cut rules (heuristics) — is composed of elementary information processes (e.g., add, subtract, compare, substitute). On the other hand, strategies may be joined in higher-level compound information processes. Strategy specifications were subsequently translated into computer programmes for simulation and testing.

The main objective of Simon was to gain better understanding of human thinking and the cognitive processes involved therein. He proclaimed that computer thinking is programmed in order to simulate human thinking, as part of an investigation aimed at understanding the latter (1). Thus, Simon did not explicitly aim to overcome the limitations of the human brain but rather simulate how the brain may work-out around those limitations to perform various tasks. His approach, followed by other researchers, was based on recording how people perform given tasks, and testing for efficacy of the process models through computer simulations. This course of research is different from the goals of novel cognitive computing.

  • We may identify multiple levels in research on cognition: an information processing level (‘mental’), a neural-functional level, and a neurophysiological level (i.e., how elements of thought emerge and take form in the brain). Moreover, researchers aim to obtain a comprehensive picture of brain structures and areas responsible for sensory, cognitive, emotional and motor phenomena, and how they inter-relate. Progress is made by incorporating methods and approaches of the neurosciences side-by-side with those of cognitive psychology and experimental psychology to establish coherent and valid links between those levels.

Simon created explicit programmes of the steps required to solve particular types of problems, though he aimed at developing also more generalised programmes that would be able to handle broader categories of problems (e.g., the General Problem Solver embodying the Means-End heuristic) and other cognitive tasks (e.g., pattern detection, rule induction) that may also be applied in problem solving. Yet, cognitive computing seeks to reach beyond explicit programming and construct guidelines for far more generalised processes that can learn and adapt to data, and handle broader families of tasks and contexts. If necessary, computers would generate their own instructions or rules for performing a task. In problem solving, computers are taught not merely how to solve a problem but how to look for a solution.

While cognitive computing can employ greater memory and computation resources than naturally available to humans, it is not truly attempted to create a fully rational system. The computer cognitive system should retain some properties of bounded rationality if only to maintain resemblance to the original human cognitive system. First, forming and selecting heuristics is an integral property of human intelligence. Second, cognitive computing systems try to exhibit common sense, which may not be entirely rational (i.e., based on good instincts and experience), and introduce effects of emotions and ethical or moral values that may alter or interfere with rational cognitive processes. Third, cognitive computing systems are allowed to err:

  • As Kelly explains in IBM’s paper, cognitive systems are probabilistic, meaning that they have the power to adapt and interpret the complexity and unpredictability of unstructured data, yet they do not “know” the answer and therefore may make mistakes in assigning the correct meaning to data and queries (e.g., IBM’s Watson misjudged a clue in the quiz game Jeopardy against two human contestants — nonetheless “he” won the competition). To reflect this characteristic, “the cognitive system assigns a confidence level to each potential insight or answer”.

Applications of cognitive computing are gradually growing in number (e.g., experimental projects with the cooperation and support of IBM on Watson). They may not be targeted directly for use by consumers at this stage, but consumers are seen as the end-beneficiaries. The users could first be professionals and service agents who help consumers in different areas. For example, applied systems in development and trial would:

  1. help medical doctors in identifying (cancer) diagnoses and advising their patients on treatment options (it is projected that such a system will “take part” in doctor-patient consultations);
  2. perform sophisticated analyses of financial markets and their instruments in real-time to guide financial advisers with investment recommendations to their clients;
  3. assist account managers or service representatives to locate and extract relevant information from a company’s knowledge base to advise a customer in a short time (CRM/customer support).

The health-advisory platform WellCafé by Welltok provides an example of application aimed at consumers: The platform guides consumers on healthy behaviours recommended for them whereby the new assistant Concierge lets them converse in natural language to get help on resources and programmes personally relevant to them as well as various health-related topics (e.g., dining options). (2)

Consider domains such as cars, tourism (vacation resorts), or real-estate (second-hand apartments and houses). Consumers may encounter tremendous information in these domains on numerous options and many attributes to consider (for cars there may also be technical detail more difficult to digest). A cognitive system has to help the consumer in studying the market environment (e.g., organising the information from sources such as company websites and professional and peer reviews [social media], detecting patterns in structured and unstructured data, screening and sorting) and learning vis-à-vis the consumer’s preferences and habits in order to prioritize and construct personally fitting recommendations. Additionally, it is noteworthy that in any of these domains visual information (e.g., photographs) could be most relevant and valuable to consumers in their decision process — visual appeal of car models, mountain or seaside holiday resorts, and apartments cannot be discarded. Cognitive computing assistants may raise very high consumer expectations.

Cognitive computing aims to mimic human cognitive processes that would be performed by intelligent computers with enhanced resources on behalf of humans. The application of capabilities of such a system would facilitate consumers or the professionals and agents that help them with decisions and other tasks — saving them time and effort (sometimes frustration), providing them well-organised information with customised recommendations for action that users would feel they  have reached themselves. Time and experience will tell how comfortably people interact and engage with the human-like intelligent assistants and how productive they indeed find them, using the cognitive assistant as the most natural thing to do.

Ron Ventura, Ph.D. (Marketing)

Notes:

1.  “Thinking by Computers”, Herbert A. Simon, 1966/2008, reprinted in Economics, Bounded Rationality and the Cognitive Revolution, Massimo Egidi and Robin Marris (eds.)[pp. 55-75], Edward Elgar.

2. The examples given above are described in IBM’s white paper by Kelly and in: “Cognitive Computing: Real-World Applications for an Emerging Technology”, Judit Lamont (Ph.D.), 1 Sept. 2015, KMWorld.com

Read Full Post »

There can hardly be a doubt that Internet users would be lost and unable to exploit the riches of information in the World Wide Web (WWW), and the Internet overall, without the aid of search engines (e.g., Google, Yahoo!, Bing). Anytime information is needed on a new concept or in an unfamiliar topic, one turns to a search engine for help. Users search for information for various purposes in different spheres of life — formal and informal education, professional work, shopping, entertainment, and others. While on some tasks the relevant piece of information can be quickly retrieved from a single source chosen from the results list, oftentimes a rushed search that relies on results in immediate sight is simply not enough.

And yet users of Web search engines, as revealed in research on their behaviour, tend to consider only results that appear on the first page (a page usually includes ten results). They may limit their search task even further by focusing on just the first “top” results that can be viewed on the screen, without scrolling down to the bottom of the first page. Users then also tend to proceed to view only a few webpages by clicking their links on the results list (usually up to five results)[1].

  • Research in this field is based mostly on analysis of query logs, but researchers also apply lab experiments and observation of users in-person while performing search tasks.     

Internet users refrain from going through results pages and stop short of exploring information sources located on subsequent pages that are nonetheless potentially relevant and helpful. It is important, however, to distinguish between search purposes, because not for every type of search looking farther than the first page is necessary and beneficial. Firstly, our interest is in a class of informational search whose purpose in general is to learn about a topic (other recognized categories are navigational search and transactional / resource search)[2]. Secondly, we may distinguish between a search for more specific information and a search for learning more broadly about a topic. The goal of a directed search is to obtain information regarding a particular fact or a list of facts (e.g., UK’s prime minister in 1973, state secretaries of the US in the 20th century). Although it is likely we could find answers to such questions from a single source (e.g., Wikipedia), found on the first page of results, it is advisable to verify the information with a couple of additional sources; but that usually would be sufficient. An undirected search, on the other hand, is aimed to learn more broadly about a topic (e.g., the life and work of architect Frank Lloyd Wright, online shopping behaviour). The latter type of search is our main focus since in this case ending a search too soon can be the more damaging and harmful to our learning or knowledge acquisition [3]. This may also be true for other types of informational search identified by Rose and Levinson, namely advice seeking and obtaining a list of sources to consult [4].

With respect to Internet users especially in the role of consumers, and to their shopping activities, a special class of topical search is associated with learning about products and services (e.g., features and attributes, goals and uses, limitations and risks, expert reviews and advice). Negative consequences of inadequate learning in this case may be salient economically or experientially to consumers (though perhaps not as serious for our knowledgebase compared with other domains of education).

The problem starts even before the stage of screening and evaluating information based on its actual content. That is, the problem is not of selectively choosing sources that appear reliable or their information seems relevant and interesting; it is neither of selectively favouring information that supports our prior beliefs and opinions (i.e., a confirmation bias). The problem has to do with the tendency of people to consider and apply only that portion of information that is put in front of them. Daniel Kahneman pointedly labeled this human propensity WYSIATI — What You See Is All There Is — in his excellent book Thinking, Fast and Slow [4]. Its roots may be traced to the availability heuristic which deals with the tendency of people to rely on the exemplars of a category presented, or ease of accessing the first category instances from one’s memory, in order to make judgements about frequency or probability of categories and events. The heuristic’s effect extends also to error in assessing size (e.g., using only the first items of a data series to assess its total size or sum). However, WYSIATI should better be viewed in the wider context of a distinction explained and elaborated by Kahneman between what he refers to as System 1 and System 2.

System 1 is intuitive and quick-to-respond whereas System 2 is more thougthful and deliberate. While System 2 is effortful, System 1 puts as little effort as possible to make a judgement or reach a conclusion. System 1 is essentially associative (i.e., it draws on quick associations that come to mind), but it consequently also tends to jump to conclusions. System 2 on the other hand is more critical and specialises in asking questions and seeking more required information (e.g., for solving a problem). WYSIATI is due to System 1 and can be particularly linked with other possible fallacies related to this system of fast thinking (e.g., representativeness, reliance on ‘low numbers’ or insufficient data). Albeit, the slow thinking System 2 is lazy — it does not hurry to intervene, and even when it is activated on the call of System 1 often enough it only attempts to follow and justify the latter’s fast conclusions [5]. We need to enforce our will in order to make our System 2 think harder and improve where necessary on poorly-based judgements made by System 1. 

Several implications of WYSIATI when using a Web search engine become apparent. It is appealing to follow a directive which says: the search results you see is all there is. It is in the power of System 1 to tell users when utilising a search engine: there is no need to look further — consider links to search hits immediately accessible on the first page, preferably seen on the screen from top of the page, perhaps scroll down to its bottom. Users should pause to ask if the information proposed is sufficient or they need to look for more input.

  • Positioning a “ruler” at the bottom of any page with page numbers and a Next button that searchers can click-through to proceed to additional pages (e.g., Google) is not helpful in this regard — such a ruler should be placed also at the top of a page to encourage or remind users to check subsequent pages, whether or not one observes all the results on a given page.

Two major issues in employing sources of information are relevance and credibility of their content. A user can take advantage of the text snippet quoted from a webpage under the hyperlinked heading of each result in order to initially assess if it is relevant enough to enter the website. It is more difficult, however, to judge the credibility of websites as information sources, and operators of search engines may not be doing enough to help their users in this respect. Lewandowski is critical of an over-reliance of search engines on popularity-oriented measures as indicators of quality or credibility to evaluate and rank websites and their webpages. He mentions: the source-domain popularity; click and visit behaviour of webpages; links to the page in other external pages, serving as recommendations; and ratings and “likes” by Internet users [6]. Popularity is not a very reliable, guaranteed indicator of quality (as known for extrinsic cues of perceived quality of products in general). A user of a search engine could be misguided in relying on the first results suggested by the engine in confident belief that they have to be the most credible. Search engines indeed use other criteria for their ranking like text-based tests (important for relevance) and freshness, but with respect to credibility or quality, the position of a webpage in the list of results could be misleading.

  • Searchers should consider on their own if the source (company, organization or other entity) is familiar and has good reputation in the relevant field, then judge the content itself. Yet, Lewandowski suggests that search engines should give priority in their ranking and positioning of results to entities that are recognized authorities appreciated for their knowledge and practice in the domain concerned [7]. (Note: It is unverified to what extent search engines indeed use this kind of appraisal as a criterion.) 

Furthermore, organic results are not immune to marketing-driven manipulations. Paid advertised links normally appear now on a side bar, at top or bottom of pages, mainly the first one, and they may also be flagged as “ads”. Thus searchers can easily distinguish them and choose how to treat them. Yet, the position of a webpage in the organic results list may be “assisted” by using techniques of search engine optimization (SEO), increasing their frequency of retrieval, for example through popular keywords or tagwords in webpage content or promotional links to the page (non-ads). Users should be careful of satisficing behaviour, relying only on early results, and be willing to look somewhat deeper into the results list on subsequent pages (e.g., at least 3-4 pages, sometimes reach page 10). Surprisingly instructive and helpful information may be found in webpages that appear on later results pages. 

  • A principal rule of information economics may serve users well: keep browsing results pages and consider links proposed until additional information seems marginally relevant and helpful and does not justify the additional time continuing to browse results. Following this criterion suggests no rule-of-thumb for the number of pages to view — in some cases it may be sufficient to consider two results pages, while in others it could be worth considering even twenty pages. 

Another aspect of search behaviour concerns the composition of queries and the transition between search queries during a session. It is important to balance sensibly and efficiently between the number of queries used and the number of results pages viewed on each search trial. Web searchers tend to compose relatively short queries, about 3-4 keywords on average in English (in German queries are 1-2 words long since German includes many composite words). Users make relatively little use of logical operators. However, users update and change queries when they run into difficulty in finding the information they seek. It becomes a problem if they get unsatisfied with a query because they could not find the needed information too shortly. Users also switch between strings of keywords and phrases in natural language. Yet updating the query (e.g., replacing or adding a word) frequently changes the results list only marginally. The answer to a directed search may be found sometimes around the corner, that is, in a webpage whose link appears on the second or third results page. And as said earlier, it is worth checking 2-3 answers or sources before moving on. Therefore, it is wise even to eye-scan the results on 2-4 pages (e.g., based on heading and snippet) before concluding that the query was not accurate or effective enough.

  • First, users of Web search engines may apply logical operators to define and focus their area of interest more precisely (as well as other criteria features of advanced search, for example time limits). Additionally, they may try the related query strings suggested by the search engine at the bottom of the first page (e.g., in Google). Users can also refer to special domain databases (e.g., news, images) shown on the top-tab. Yahoo! Search, furthermore, offers on the first page a range of results types from different databases mixed with general Web results. And Google suggests references to academic articles from its Google Scholar database for “academic” queries.

The way Interent users perceive their own experience with search engines can be revealling. In a survey of Pew Research Center on Internet & American Life in 2012, 56% of respondents (adults) expressed strong confidence in their ability to find the information they need by using the service of a search engine and an additional 37% said they were somewhat confident. Also, 29% said they are always able to find the information looked for and 62% said they can find it most of the time, making together a vast majority of 91%. Additionally, American respondents were mostly satisfied with information found, saying that it was accurate and trustworthy (73%), and thought that relevance and quality of results improved over time (50%).

Internet users appear to set themselves modest information goals and become satisfied with the information they gathered, suspectedly too quickly. They may not appreciate enough the possibilties and scope of information that search engines can lead them to, or simply be over-confident in their search skills. As suggested above, a WYSIATI approach could drive searchers of the Web to end their search too soon. They need to make the effort, willingly, to overcome this tendency as the task demands, getting System 2 at work. 

Ron Ventura, Ph.D. (Marketing)

Notes:

(1) As cited by Dirk Lewandowski (2008), Search Engine User Behaviour: How Can Users Be Guided to Quality Content, Information Service & Use, 28, pp. 261-268 http://eprints.rclis.org/16078/1/ISU2008.pdf ; also see for example research by Bernard J. Jansen and Amanda Spink (2006) on How Are We Searching the World Wide Web.

2) Daniel E. Rose & Danny Levinson (2004), Understanding User Goals in Web Search, ACM WWW Conference, http://facweb.cs.depaul.edu/mobasher/classes/csc575/papers/www04–rose.pdf 

(3) Dirk Lewandowski (2012), Credibility in Web Search Engines, In Online Credibility and Digital Ethos: Evaluating Computer-Mediated Communication, S. Apostel & M. Fold (Eds.) Hershey, PA: IGI Global (viewed at: http://arxiv.org/ftp/arxiv/papers/1208/1208.1011.pdf, 8 July ’14)

(4) Daniel Kahneman (2011), Thinking, Fast and Slow, Penguin Books.

(5)  Ibid. 4.

(6) Ibid. 3

(7) Ibid 1 (Lewandowski 2008). 

 

         

Read Full Post »