Saturday, May 30, 2015

Decision making :one of multiple category options

In the previous post, Decision-making learning from one’s mistakes., I provided evidence that selective attention to items that were retrieved into working memory were a major factor in making good decisions. This has generally unrecognized educational significance. Rarely is instructional material packaged with foreknowledge of how it can be optimized in terms of reducing the working memory cognitive load. New research from a cognitive neuroscience group in the U.K. is demonstrating the particular importance this has for learning how to correctly categorize new learning material. They show that learning is more effective when the instruction is optimized ("idealized" in their terminology).

Decisions often require categorizing novel stimuli, such as normal/abnormal, friend/foe, helpful/harmful, right/wrong or even assignment to one of multiple category options. Teaching students how to make correct category assignments is typically based on showing them examples for each category. Categorization issues routinely arise when learning is tested. For example, the common multiple-choice testing in schools requires that a decision be made on each potential answer as right or wrong.

In reviewing the literature on optimizing training, these investigators found reports that one approach that works is to present training in a specific order. For example, in teaching students how to classify by category, people perform better when a number of examples from one category are presented together followed by a number of contrasting examples from the other category. Other ordering manipulations are learned better if simple, unambiguous cases in either category are presented together early in training, while the harder, more confusing cases are presented afterwards. Such training strengthens the contrast between the two categories.

The British group has focused on the role of working memory in learning. Their idea is that ambiguity during learning is a problem. In real-world situations that require correct category identification, naturally occurring ambiguities make correct decisions difficult. Think of these ambiguities as cognitive "noise" that interferes with the training that is recalled into working memory. This noise clutters the encoding during learning and clutters the thinking process and impairs the rigorous thought processes that may be needed to make a correct distinction. In the real world of youngsters in school, other major cognitive noise sources are the task-irrelevant stimuli that come from multi-tasking habits so common in today's students.

The theory is that when performing a learned task, the student recalls what has been taught into working memory. Working memory has very limited capacity, so any "noise" associated with the initial learning may be incompletely encoded and the remembered noise may also complicate the thinking required to perform correctly. Thus, simplifying learning material should reduce remembered ambiguities, lower the working memory load, and enable better reasoning and test performance.

One example of optimizing learning is the study by Hornsby and Love (2014) who applied the concept to training people with no prior medical training to decide whether a given mammogram was normal or cancerous. They hypothesized that learning would be more efficient if students were trained on mammograms that were easily identified as normal or cancerous, and did not include examples where the distinction was not so obvious. The underlying premise is that decision-making involves recalling past remembered examples into working memory and accumulating the evidence for the appropriate category.  If the remembered items are noisy (i.e. ambiguous) the noise also accumulates and makes the decision more difficult. Thus, learners will have more difficulty if they are trained on examples across the whole range of possibilities from clearly evident to obscure than if they were separately trained on examples that were clearly evident as belong into one category or another.

Initially a group of learners was trained on a full-range mixture of mammograms so the images could be classified by diagnostic difficulty as easy or hard or in between. On each trial, three mammograms were shown: the left image was normal, the right was cancerous, and the middle was the test item requiring a diagnosis of whether it was normal or cancerous.

In the actual experiment, one student group was trained to classify a representative set of easy, medium, and hard images, while the other group was trained only on easy samples. During training trials, learners looked at the three mammograms, stated their diagnosis for the middle image, and were then given feedback as to whether they were right or wrong. After completing all 324 training trials, participants completed 18 test trials, which consisted of three previously unseen easy, medium and hard items from each category displayed in a random order. Test trials followed the same procedure as training trials.

When both groups were tested on samples across the range in both conditions, the optimized group was better able to distinguish normal from cancerous mammograms in both the easy and medium images. Note that the optimized group was not trained on medium images. However, no advantage was found in the case of hard test items; both groups made many errors on the hard cases, and optimized training yielded poorer results than regular training. 

We need to explain why this strategy does not seem to work on hard cases. I suspect that in easy and medium cases, not much understanding is required. It is just a matter of pattern recognition, made easier because the training was more straightforward and less ambiguous. The learner is just making casual visual associations. For hard cases, a learner must know and understand the criteria needed to make distinctions. The subtle differences go unrealized if diagnostic criteria are not made explicit in the training. In actual medical practice, many mammograms actually cannot be distinguished by visual inspection—they really are hard. Other diagnostic tests are needed.

The basic premise of such research is that learning objects or task should be pared down to the basics, eliminating extraneous and ambiguous information, which constitute “noise” that confounds the ability to make correct categorizations.

In common learning situations, a major source of noise is extraneous information, such as marginally relevant detail. Reducing this noise is achieved by focus on the underlying principle. Actually I stumbled on this basic premise of simplification over 50 years ago when I was a student trying to optimize my own learning. What I realized was the importance of homing in on the basic principle of what I was trying to learn from instructional material. If I understood a principle, I could use that understanding to think through to many of the implications and applications.

In other words, the principle is: "don't memorize any more than you have to." Use the principles as a way to figure out what was not memorized. Once core principles are understood, much of the basic information can be deduced or easily learned. This is akin to the standard practice of moving from the general to the specific. Even so, general ideas should emphasize principles.

Textbooks are sometimes quite poor in this regard. Too many texts have so much ancillary information in them that they should be thought of as reference books. That is why I have found a good market for my college-level neuroscience electronic textbook, “Core Ideas in Neuroscience,” in which each 2-3 page chapter is based entirely on each of the 75 core principles that cover the broad span of membrane biochemistry to human cognition.. A typical neuroscience textbook by other authors can run up to 1,500 pages.