Sunday, September 04, 2011

The Transfer Challenge to Expertise

Consider this relatively commonsensical view of problem solving:

“We absorb information in some common generic way and then apply our individual talents to using that information to solve problems.”

As straightforward and intuitive as this description sounds, it contains a counter-productive assumption about knowledge, confusing it with information. We tend to assume that the things we experience are experienced in the same way by other people. We might all look at the same world, but we make different observations and draw different conclusions from it. The knowledge base we each build over our lifetime is not just a straightforward result of the information presented to us; it is also in part a result of how we organize that information, which can be very different from one person to another. This is the foundation of the expertise model.

Organization Matters

It is not enough to have the right information to solve a problem, the information must also be organized in a way that lets us think about it in the right way.

Expertise is a key factor in effective problem solving because it permits us to organize information about a domain, recognize patterns in that domain, apply domain knowledge to new kinds of problems, and incorporate new information about that domain.

The expertise model gives us several key insights into the problem solving process that the commonsensical view above misses:

It tells us that the way information is organized is critical to how that information can be used for problem solving. Many problem solving principles will deal with how information is best organized to solve problems.
It tells us that in certain key areas, deep accurate understanding is important and not just superficial familiarity. Both intuitive decision making and more formal methods rely on deep accurate understanding acquired from experience.
It tells us that a great deal of deliberate practice with good feedback is needed to acquire deep understanding.

It tells us that more skillful problem solving emphasizes principles, while less skillful problem solving relies on procedures.
In spite of the tremendous power of the expertise model, it has its own limitations as well. Expertise lets us detect meaningful patterns of information in particular domains because of the way we organize our own knowledge.

The organization of domain-specific knowledge in our mind has important implications. It means that experts have differently organized knowledge for thinking in different domains. The fact that the expertise model is so thoroughly domain-specific forces us to now confront the most serious challenge of all to the expertise model: the challenge of transfer. If it takes so much deliberate practice to become good in a given domain, how does anyone manage to become good at more than a very narrow range of activities? How do our narrowly cultivated abilities support other activities? Or do we have important abilities that are not domain-specific as well? What does it take to apply our hard-won expertise to problems different than the ones we specifically practiced for?

The Failure of Mental Exercise

At one time, it was widely believed that people could develop their mind by doing mental work such as solving logic puzzles, learning mathematics, reading classics of literature, and learning to speak Latin. A long history of sometimes large scale research on this approach to mental ability revealed it to apparently have very little promise.[1] Literacy in general, while valuable for its own sake, simply does not have much effect on other thinking abilities.[2] Our ability to solve puzzles doesn’t tend to generalize very well unless the training specifically teaches the underlying patterns and provides us with a way of remembering them.[3] We don’t automatically apply the lessons of solving one problem to solving a structurally similar but different problem. The different appearance of problems tends to throw us off. When we learn strategies for solving problems, we tend to learn them in a way that is tied to the specific kinds of problems that we used for learning them. Expertise, the research confirms, tends to be very context specific.

How Transfer Does Happen: Two Roads

The problem with this result is that while it seems consistent with the expertise model, it doesn’t quite make sense in terms of our everyday experience. All of us routinely do apply what we know to new kinds of problems. We aren’t equally incompetent in dealing with any sort of novel problem; our existing expertise clearly does sometimes give us an advantage in another domain. We also see negative influence of expertise when our existing abilities interfere with our attempts to perform in a similar domain. Expertise does seem to transfer between domains under some conditions. The question is what those conditions might be.

Research done in the late 1980’s and early 1990’s confirmed that transfer of ability between domains does occur consistently under certain conditions. One cognitive psychologist working with preschool children on simple tasks discovered that the 3 and 4 year olds could use lessons they learned under one set of conditions in a completely different set of conditions, but especially if they were shown how the different problems resembled each other and how the goals were similar, they were familiar with the problem areas, the examples also had rules associated that the children figured out for themselves, and if the learning took place in a social context that specifically encouraged them to spell out the principles, explanations, and justifications to use.[4]

Research such as this led to a general two-pronged theory of transfer, proposed by David Perkins and Gavriel Salomon. The theory is based on the finding that transfer sometimes takes place between similar domains, and sometimes takes place between very dissimilar domains, and that these seem to happen under different conditions.[5]

What the Perkins and Salomon theory calls “low road transfer” happens when situations appear to us to be similar according to simple perceptual cues rather than any deep structural pattern. This seems to be a matter of stimulus triggers. Specific elements in the situation help us recognize and apply skills and knowledge from our memory based on recognizing those elements from our practice. Since low road transfer is pattern-bound, it doesn’t generally lead to transfer to different situations. Practicing under a variety of different conditions however can help is gradually stretch our skills from one context to a similar one to generalize our skills further. Low road transfer is a result of the variety of conditions under which we practice rather than any specific cognitive skills or strategies aside from those specific to the domain. Low road transfer is a perceptual-memory phenomenon.

When a situation bears a superficial similarity to one we’ve trained for, we recognize stimulus patterns and our expertise is evoked via low road transfer. This is how many people manage to drive a truck reasonably well after having learned to drive a car for example. Even though the mechanics are very different, the steering, pedals, and so on are all familiar enough to trigger our learned skills for driving. That is, until we find ourselves in a situation where the fit isn’t so good between our skills and the ones that are needed.

What the Perkins and Salomon theory calls “high road transfer” seems to be a completely different matter. High road transfer involves the deliberate and mindful abstraction of principles during practice and using more general cognitive skills and strategies to apply them to completely different situations. In high road transfer, the learner actively seeks connections between different situations in which to apply the principles they’ve learned. High road transfer is a cognitive phenomenon.

We see that the similarity mechanism of transfer is limited. It only works for relatively similar situations and it works in a very automatic and unthinking way. To apply expertise to a very differently appearing situation with underlying structural similarity (such as we might need for more abstract problem solving) we need high road transfer and we need to use abilities we associate with conscious reflection. This allows us to transfer expertise from deliberate abstraction of principles to entirely different kinds of problems.

The lesson of the transfer challenge to expertise is that expertise does not automatically apply outside of its domain. We have to very deliberately either: (1) work on practicing in widening ranges of situations to facilitate generalization or (2) work on abstracting and applying general principles mindfully from our practice, or both.

Conclusion: Transfer and Expertise

We’ve seen that expertise is a very powerful model that explains in some detail how we organize tacit knowledge for recognizing patterns and solving a particular domain of problems. This appears to explain the lion’s share of differences in human abilities in problem solving. We’ve also seen that expertise can be acquired in such a way that it can be generalized to an increasingly wider range of conditions and in a way that makes it less likely to fail catastrophically under extreme conditions, making expertise a potentially very robust resource for problem solving.

We’ve also seen that the expertise model misses a small but critical aspect of problem solving; it does not tell us how people manage to deal with surprises or with domains that are characterized by surprises. The expertise research consistently shows strong dependence on specific contexts. We do not automatically generalize our skills or strategies to new kinds of problems just by acquiring deep expertise in a domain.

Novelty offers our most serious challenge to the power of expertise. The very concept of expertise implies domain-specificity, and domain-specificity implies that expertise is honed to deal effectively with a particular range of situations. Novelty, both within a domain and outside that domain, creates problems for the standard expertise model that need to be addressed.

Novel but superficially similar situations can be handled through expertise, but only if we specifically widen our practice to deal with a broader range of conditions.

Completely novel situations in other domains that don’t resemble the ones we practice for except in terms of their underlying deep structure can be handled through expertise as well, but only by deliberate attention to learning and applying general principles as well as acquiring domain expertise.

The Story So Far: Going Beyond Expertise

The expertise model tells us how we acquire useful patterns of tacit knowledge from experience through deliberate practice with good feedback. The expertise model explains how we deal effectively with the sorts of situations where we have accumulated extensive practice. Expertise thus acquired becomes part of our intuitive understanding of situations, enhancing, modifying, and extending our existing commonsense intuitions.

The expertise model also challenges us to explain how it is that we are able to deal with extreme yet realistic conditions and novel problems even though expertise tends to be very context-specific. Applying expertise to very different situations requires deliberate mindful work at abstracting principles and applying them through our capacity for reflective thinking. This kind of reflective thinking is not adequately captured by the expertise model. Either we need to expand the expertise model to handle the challenges we have identified, or else we need a more expansive concept to describe our abilities.

[1] (Thorndike & Woodworth, 1901), (Thorndike, The influence of first year Latin upoin the ability to read English, 1923)

[2] (Scribner & Cole, 1981)

[3] (Simon & Hayes, Psychological differences among problem isomorphs, 1977)

[4] (Brown, 1989)

[5] (Perkins & Salomon, 1987), (Perkins & Salomon, Teaching for transfer, 1988), (Salomon & Perkins, 1989)

Saturday, September 03, 2011

The Unpredictability Challenge to Expertise

We almost universally recognize the legitimacy of experts in a number of different domains. In many academic fields such as mathematics, sciences, history, literature, and other academic areas, some people know much more and consistently perform much better than others in tests of ability. Similarly for many professional fields and various sports and games, we recognize that there are experts who outperform the majority of us.

A key finding in modern learning and human performance research has been the discovery of how expertise is acquired.[1] This discovery became possible with the advent of cognitive science, allowing us to model the human brain as an organized collection of information rather than just a collection of behavioral patterns. As we learned about the specific differences between novices and experts[2] in each field, we discovered certain general principles that apply to a very wide range of different fields.

A formidable body of this type of research has overturned the intuitive view that novices and experts differ because some people are simply more naturally talented than others. Experts seem to perform so effortlessly that we tend to attribute great natural ability to them rather than a different kind and degree of experience. Contrary to this intuition, expertise via deliberate practice is our best model so far of individual differences in ability in a wide range of activities. Expertise is a result of experience, and not just any experience, but deliberate practice where we meet challenges in that domain, are immersed in purposeful practice, gain knowledge from other people who are already good at it, have good coaches, and benefit from quality feedback for our performance.[3]

At the same time as verifying the legitimacy of expertise in many fields and establishing the central role of deliberate practice, social environment, coaching, and feedback, we have also discovered that there are some fields where our performance doesn’t benefit from these factors.

In spite of the tremendous power of the expertise model, it has its own limitations as well. Expertise lets us detect meaningful patterns of information in particular domains because of the way we organize our own knowledge. This assumes that there are meaningful patterns to detect that human beings are capable of using effectively. This is not always the case.

Where the Experts Fail

In the 1950’s, research into medical diagnosis and prognosis revealed something shocking: presumed experts didn’t seem to predict medical outcomes any better than novices. This line of research continued over time to demonstrate that prognosis and diagnosis in clinical work in medicine was often not improved by experience when experts relied upon informal gathering of data and their trained intuitions.[4]

Professional experience and presumed expertise also seem to make no difference in predicting the outcome of psychotherapy by psychologists, who it turns out also fare no better than less trained individuals.[5]

If there is a skill to predicting medical and psychotherapeutic outcomes in general, it doesn’t seem to be acquired from the standard professional training, or typically through experience with patients, and it isn’t obvious how else it might be acquired. There is perhaps good reason why some experienced doctors seem very reluctant to make predictions about outcomes, and maybe more of them should heed this lesson.

As a result of the difficulty of prediction in areas like this, novices using simple formal statistical methods have often outperformed the experts in tests in spite of the greater experience and training of the experts (or perhaps in some cases partly because of it).

This is not by any means to imply that statistical methods are always superior to expertise, even in a particular field where clinical experience has proven less than optimal. However, it does give us good reason to pause and reflect on the meaning of this finding for the expertise model. In some domains, the best we can do for prediction is provided by a simple statistical method; and the value of expertise in particular reaches a limit fairly early on in the training for those fields.

What Makes the Value of Expertise Vary So Much?

Research into the value of expertise in different domains shows it to vary[6] with:

1.the level of inference[7] required (moderate levels of inference are more conducive to using expertise than high levels of inference),
2.whether experience or training available is adequate to confer expertise,
3.whether the conditions and instruments available allow for the expression of expertise

What this tells us is that even though expertise helps us make sense of complex situations by recognizing patterns, there is also a limit to how well acquired expertise can help us make better judgments in very complex situations. The more specialized knowledge we need in a field just to understand what is going on, the more likely it is that expertise will fail us when the situation requires a great deal of challenging inference. In the most complex fields at least, it may be that intelligence can also play an important role alongside expertise.[8]

Both intelligence and expertise play some role in every field, but each is more important to some fields than others and at different points in the development and expression of ability. The relevance of intelligence in a field seems to depend to a large degree on the role that abstract reasoning plays in success in that field. The relevance of expertise is more general. The role of expertise in a field depends on how well the situation is made comprehensible to the expert through specialized tools, the quality of their training and experience, and the kinds of conditions in which they have to perform.

Even allowing for a role for intelligence in particularly difficult technical fields requiring very high inference levels, there are fields where neither expertise nor intelligence nor any combination of the two seems to predict performance any better than simple methods.

We’ve discovered that in some areas, experts perform significantly better than non-experts and consistently outperform computer models of various kinds because of their rich background of task-relevant skills and knowledge. In .other areas, simple computer models, statistical indexes, and non-experts consistently outperform experts.

Intelligence may play more of a role in ability in highly technical domains where a high level of inference is often required in addition to recognizing important patterns. Domains are apparently not all equal with regard to what it takes to be good at them.

How Surprises Can Negate Expertise

The difference has to do with the varying role of understanding the situation for solving problems in different fields, and the role that surprise plays in each field. Fields involving things that move freely and things that scale wildly rather than behaving according to standard statistical methods[9] tend to produce surprises that can’t be managed primarily by either intelligence or expertise or both. In these areas the requisite intelligence is relatively low and the practical role of expertise is relatively marginal because reasoning doesn’t help much and it is particularly difficult to get the necessary skills even if you can identify them. So in these fields, simple statistical rules can sometimes perform as well as any expert, regardless of their IQ.

For examples of fields more or less dominated by surprises think of stockbrokers, risk management advisors, clinical psychologists, counselors, psychiatrists, admissions officers, court judges, economists, financial advisors, and intelligence analysts. Think in general of all the fields where experts fare poorly compared to non-experts, where overconfidence cancels out the benefits of expertise, or where time spent in formal practice has relatively little impact on effective outcomes.

In these fields, formal domain-specific expertise and general intelligence provide relatively little advantage in producing good outcomes compared to simple algorithms, direct local observation, direct experience, practical skills, and domain general problem solving skills. Formal expertise and intelligence in these fields especially tends to produce overconfidence more than real predictive ability. It’s not impossible that there may be some real experts in these fields who fare better than others, but they are particularly difficult to identify and train with formal methods.

Not all fields are dominated by surprises regardless of intelligence and expertise. Think of fields involving things that stay put or else move within strictly defined ranges according to physical laws or arithmetic or statistical relationships. These are much better suited to intelligence and domain-specific expertise because in those fields a better understanding of identifiable patterns and potentially complex information does tend to lead to better prediction of outcomes. Think of theoretical mathematicians and physicists, astronomers, test pilots, firefighters, livestock, grain, and soil judges, accountants, chess masters, insurance analysts (who deal with Gaussian topics like mortality), competitive athletes, and surgeons. Think in general of the many fields studied by expertise researchers where deliberate formal practice yields measureable improvements in results, and where the critical skills can be identified and trained.

We Don’t Learn Well from History

Part of the problem with expertise in fields where surprise plays an important role is that we don’t learn well from history in general. One of our consistent biases is that we systematically overweight the likelihood of events that actually happened, relative to ones that didn’t happen (but could have). This means that we have a very strong predisposition to describe events that happened as if they were fated to happen that way. This also means that we tend to think of our descriptive stories as if they were also explanations, not just descriptions. The remarkable power of stories becomes a disadvantage for explanation because the narrative content tends to replace our ability to analyze cause and effect.[10]

We become experts by being exposed to similar conditions over and over again and learning from consistent patterns in our experience. When a domain is characterized by events that are relatively uncommon yet influential, our confidence in our ability to predict events in that domain tends to grow way out of proportion to our actual ability to predict or explain the course of events. Our hindsight bias (“I knew it all along”) often kicks in to replace our missing explanatory ability.[11]

The human mind is particularly well suited to remembering and making sense of events after the fact by weaving facts into a plausible narrative, and particularly poorly suited to capturing actual frequencies of events in order to use that information in other judgments. Our common sense excels at generating plausible stories for what happens, our expertise then generates trained intuitions that add to our confidence in our explanations, but in some cases does not also add to our explanatory ability. Then history leaves us with only a single chain of events to explain, the one that actually happened. We infer from all of this that we are explaining why a sequence of events took place in a particular situation, whereas we have often only described the events, not explained them.[12]

Conclusion: Surprises and Expertise

We saw in the previous section that extremes of arousal can negate some kinds of expertise, especially expertise relying on fine motor skills. We also saw that our mindset can determine whether expert performance is retained during high arousal or fails catastrophically. In addition we saw that our ability to flexibly adapt our responses to novelty in the situation is hampered by high arousal.

Now we see that novelty offers a more general and more serious kind of challenge than just our tendency to lock in to central stimuli under high arousal. When relatively uncommon events tend to be influential in a domain, the power of expertise to help us predict and explain events is severely compromised and often even negated entirely. In these cases we have a compelling natural tendency to tell plausible stories and rely on them as explanations, and additional expertise only serves to increase our overconfidence.

[1] A June 2008 review of major trends in expertise research: (Charness & Tuffiash, 2008)

[2] An expert is typically defined for research purposes as someone who consistently performs more than two standard deviations above the mean average performance on representative tasks for their domain, assuming that ability in the field can be represented by measureable tasks and that this ability is normally distributed (Ericsson & Charness, 1994). Experts defined in this way are assumed to be roughly the top 5% of the performers in a field.

[3] The body of research has been variously summarized in popular books by journalists but a far better source for reviewing the evidence directly is the edited technical article collection: The Cambridge Handbook of Expertise and Expert Performance (Ericsson, Charness, Feltovich, & Hoffman, 2006)

[4] Classic early research showing the limits of human judgment from experience was done by Paul E. Meehl. Meehl demonstrated the limits of informal aggregation of data and prognostication by presumed experts in clinical situations such as diagnosing patients and predicting medical outcomes (Meehl, 1954). An influential review of research showing the superiority of actuarial vs. clinical judgment appeared in the journal Science in 1989: (Dawes, Faust, & Meehl, 1989)

[5] (Dawes, 1994)

[6] (Westen & Weinberger, 2005)

[7] The level of inference means the amount of specialized individual knowledge needed to understand what is going on. Situations with low levels of inference are understandable by most people, those with high levels of inference are only accurately understood by experts. Even experts utilize their abilities better in situations of lower levels of inference.

[8] There is a lot of ongoing controversy about various aspects of intelligence measurement and what it can tell us, but one of the things that most theorists agree on regarding individual differences in intelligence measurements is that they seem to correspond in some sense to our capacity to handle complexity. (Neisser, et al., 1996)

[9] This was one of the main points made by Nassim Nicholas Taleb in his entertainingly and ironically sharp book exhorting the importance of epistemological humility in the face of this sort of unpredictability in important domains, The Black Swan (Taleb, 2007)

[10] For examples in technical literature making this argument more clearly, see: (Lombrozo, 2006), and (Lombrozo, 2007). The point is made even more emphatically in (Dawes R. , 1979). Formal techniques for causal analysis take the lure of stories explicitly into account by using various methods to compensate for it and force analysts to think in causal terms rather than relying on our more natural instincts for telling stories about what happened. (Gano, 2008)

[11] A good technical article introducing hindsight bias and the related idea of “creeping determinism” (what happened is what was most likely to happen) is (Fischhoff, 1982)

[12]There is a more detailed discussion of the difference between stories and explanations in the chapter History is a Fickle Teacher in (Watts, 2011, pp. 108-134)