edited by DAVID W. BROOKS
educational research
118 Henzlik University of Nebraska Lincoln, NE 68598-0355
Connectionism David Fowler and David W. Brooks 118 Henzlik, University of Nebraska, Lincoln, NE 68508-0355 Connectionism and a Focus on Neurons It is difficult to imagine the world as it is presented to a newborn organism of a different species, no less our own. Indeed,the conventionsofsociety,the remembrances ofsensory experience, and, particularly, a scientific education make it difficult to accept the nation that the environmentpresented to such an organism is inherently ambiguous: even to animals eventually capable of speech such as ourselves, the world is initially an unlabeled place. Gerald Edelman ( I ) uses these words to introduce readers to his book Neural Darwinism. Edelman's work deals with a new theory of brain function called connectionism. Chemistry teachers strive to bring about student learning. The better a teacher understands the leamine process, the more successful that teacher will be. For example, dostudents learn better by just listening attentively or by participating actively? Connectionism is too new and unrefined to have served a s the basis for research on instruction or to have helped answer such questions. However, connectionism frames explanations for most of what is known about instruction. Descri~tionsof learnine at the cellular and even molecular leveiare being refined by biology and neuroscience (2). Connectionism locates the intellectual Drocesses of an organism within its nervous system. It ascribes mental phenomena to both temporary and lasting changes a t the synaptic level. This continuously self-modifying nervous system controls behavior and thus subsequent learning. "In a fundamental sense, the connections of a network are both hardware and software" (3). Memory means something within the brain of an organism that permits behavioral outputs (speech, motion) which re-create past inputs (perceptions) or t h a t demonstrates a continuing awareness of them (4).Remember means to use memory. Distributed means that information processing involves many functional regions of the brain and millions of cells within each region (5).The brain does not contain an analog of the central processing unit found in conventional (von Neumann) computers. There is no "central executive" in the brain, not even a group of linked executives. Parallel means that the various neuroAbout This Column The Editor has requested an occasional series of papers about2'whatresearch says to the chemistryteacher". Since teaching implies learning, the first atticle is about learning. One other article on teacher-led research has been scheduled. Prospective authors should submit manuscripts to D. Brooks at the above address.
748
Journal of Chemical Education
nal activities are concurrent: Serial operations occur less oRen than in a conventional computer. Although opportunities for linkmg and feedback (or reentry) abound, independent concurrent operations are hallmarks of this model (6).The brain modifies itself through an exquisitely subtle system of neurotransmitters: Sensory inputs modify memory; even the act of remembering modifies memory. Different regions of the brain are specialized (7):Vision memory is in one region; motor memory, in another. In connectionism the brain is a distributed, parallel, self-modifying system with content-addressable, associative memory (8).Understanding connectionism will lead to improved learning through improved teaching. In a neuronal model, examples of initiating events in a sensory input group are definite and discrete: a photon striking a retina; a suitably shaped molecule binding at a surface site of an olfactory bulb. The chemical mechanisms that transform such events into the firing of a neuron vary among sensory input systems. One or several events may initiate the firing. When a group of related (nearby) neurons fires synchronously, they may in turn fire or inhibit another neuron removed from the primary event. The secondary group of neurons can be arranged or connected to respond to features (parts of shapes, for example). The secondary group may fire into another group, which may fire to still another group, and so forth. Groups of cells may respond as feature detectors in a primitive, "close to the event" fashion (9).In fact, the information received by the cerebral cortex is already highly processed. At the other end of the system, outputs involve actual or potential muscle movements. You are probably not speaking at this moment, but you immediately could speak and utter either the words on the page or your thoughts. Muscle movements are initiated by neuronal firing, and every muscle fiber is controlled by a single motor neuron that may control other fibers (10). Behind a primary output group of motor neurons there is a secondary group of neurons, and so forth. By conventional wisdom, between input and output groups there are central decisionmakers or executives. The connectionist model uses no centralized system. Information, ceaselessly moving forward from input to output, is changed in a connectionist system as it is passed along so that receiving groups do not know if or where modifications were introduced (11).Although we perceive a world that sends holistic perceptions, the perception channels work independently. When related information from two channels is blended, modifications may be made by one or both channels. Optical illusions are a well-known result. There is no way to go back and check the information to see if part was changed in processing. In connectionism groups of
neurons respond to synchronies among outputs of different sensory systems. Thus, the left hand does know what the right hand is doing. While information is being processed, changes may occur in the entire svstem to make synapse firings easier or more difficult. When a synapse f&, itreleasesneurotransmitter molecules into the synaptic cleft. The . postsynaptic . neuron may then fire, thus &ing its synapses (at their presynaptic sides). The structure of the synapses changes as the system functions (12). Genetics plays a n important and very well-documented role in learning (13). However, the human genome is far too small to include enough information to describe all the synaptic connections in the brain. Furthermore, we have no microelectrician, homunculus, to do the wiring for us (14). The dynamics of synapses are lifelong, wntinuously ongoing processes. There are several big questions to ask. Can a connectionist system with no centralized, preordained, decision-making executive work elsewhere? The answer is a definite "yes". Many computer models, called neural networks, demonstrate this kind of behavior (15). Is there consistent experimental support for this model? Many animal studies support this model, with the most powerful using slugs (Aplysia californica or Hermissenda cmssicornis, for example). They have few, but large, neurons and are thus easy to study Although slugs are not robust (they don't solve gas law problems), they can be trained (16). Why would a connectionist system evolve? Since the systemis parallel, it is very fast. Humans realize extremely quickly when they know something, and just a little less quickly when they don't. (Haveyou ever been to Fairbanks, Alaska?) Loss of even a relatively large amount of tissue (neurons) leads to "graceful degradation" rather than to a dvsfunctional svstem (17). Perhaus the most important ~" advantage of a connectionist system is its mode of Ehange. To retroactivelv chanee somethine in a central filine svstem, you must" pull 21pertinent-documents and cLaige them one at a time. This is not true in a connectionist system. For better or worse, previously stored features are changed by cnrrent learning. If an input completely or partially mimics a previously experienced input, it goes to the same neuronal groups. Suppose an early experience was consolidated (the system changed to remember it). Unless aging, disease, time, or subsequent learning change the synapses, they remain in place to respond to current events as to the original event. The content of the current message sends it to the same place in the neuronal system as the last message. Kohonen called this a content-addressable system (18). Apartial input activating only a few groups may cause an entire collection of groups to fne. (You look through the bakery store window, see the pumpkin pie, and somehow "smell" it too.) As the eroups activate, the portions of the stored information that are not part of th; current input message also are activated: The connectionist model is associotimn. -. ....-.....
A connectionist system is distributed, parallel, and selfmodifying with content-addressable, associative memory. Once a parallel, distributed system is started, it can learn. How does such a complex svstem eet started? The readers of this Journal teach learners who are usually 15-25 vears old and have connectionist svstems that are already well-established. Brain hemispheres first appear on thc 20th dav of eestation (7,.Infants form neural connections in the"wo1;b: They are born kicking and screaming. The process of making and refining connections, called deuelopment, continues throughout life. Most dynamics occur early in life. Infants lose a large fraction (about half)
-
of some of their neurons during their first two years. Neurons that do not group will die and be resorbed. Learning to eye-track, grasp, walk, and talk represent early stages of connecting of a neuronal system. Learning is exactly what a neuronal system has evolved to accomplish. People must learn how to produce an output. When our studentslearnedto walk, they fell, withthe past experience modifying the next attempt.As teachers, we are involved in this process. Our students respond to inputs from us in the same way. Our inputs shape their responses. There must be some output to begin the process. Learning involves tyingoutputs to inputs. The learning that chemistry teachers bring about requires feedback related to performance. When "the light goes on", that is, when a desired neuronal output emerges, an organism can enhance the likelihood of remembering the experience. A two-step mechanism has been proposed for this process (19).First an electrical pulse forces maenesium ions from the ion channels they block in a cell neuronal membrane. Asecond pulse, abouti00 msec later, permits calcium influx through those same channels. Calcium facilitates lasting synaptic changes. The basic precepts ofthis model are established. Concrete concepts (calcium channel, neurotransmitter, N-methyl-Daspartate receptor, synapse) replace abstract concepts (mind, consciousness, working memory). The model is still extremely dynamic,however, with exciting new experimental results being reported weekly Neuroscience has become one of the most fascinating areas of current research. Applications Much is known about effective teaching. In the future, connectionism should help frame explanations for most of what is known about instruction, just as it should explain existing data about perception and other mental phenomena. Research on learning may become an unfamiliar blend of assessing performance on cognitive tasks and of measuring physiological properties, such as response times and eye movements. Many psychologists already observe eye movement in studies of reading. Positron emission tomography already has beenused as a toolinstudies of wgnition (20).Below we attempt to use the notions of connectionism to make several zeroth-order analyses of everyday issues related to teaching and learning. The More Content, the Better Connectionist systems are content-addressable. They work better when potential addresses contain memories of related, previous perceptions. The more previous exposure students have to science, the better. Some collepe teachers want high schools to send them ..empty vessels", students with good basic skills hut no prrconcrptims about science. This is not our description of n good science student. Csing terms, exploring phenomena, making predictions, and performine exoeriments are iust the kinds of ~recolleeeexoeriences a student needs. (Vocabulary learned without context and memorized like nonsense svllables in ~svcholow experiments will not he productwe.kthough theaddrcss i6 not empty, it is still not associated with othcr addresses.~
- .
- .
Experience, Not Age, Is What Counts Connectionism includes no preconditions about time or development. The authors first started exploring connectionist models from the perspective of neb-~iaGtians when they used stage models. Common observations typical with Piagetian tasks wncern age: Young children do not "conserve water" in a pouring experiment while older children do; older high school students use proportional reasoning while junior high school students do not. A conclu-
Volume 68 Number 9 September 1991
749
sion usuallv drawn is that the students are a t a develoomental stage appropriate to learning the task. In the connectionist model. it is not clear whv such limits exist. An associative connectionist system explains why learners appear to go through stages. A time factor is needed. Developmental delays measured in days might be justified, but not years. Wollman and Chen (21)provided a successful example of teaching formal operations "prematurely". Anton Lawson recently reported success teaching second grade students material oreviouslv considered bevond their level (22). . . We now expect to see more such reports using instruction that demands frequent learner output, provides frequent feedback, and induces learners to use both the content-addressinn and associative features of their connectionist learnine systems.
-
-
Hands-on "Hands-on" experiences are not obligatory when dealing with either the contentaddressable or the associative nature of a connectionist system. Hands-on experiences are useful as far as thev facilitate wntent-addressing or associations. When compared to alternatives, they may be the best way to engage a connectionist system. In hands-on instruction, there is more output and related feedback. whether eenerated bv a teacher or a Deer. Thus. hands-on situations often yieli inherently betier instruc: tion. beine better suited to content-addressine and developing associations.
-
-
Modifying Associations Connectionist systems are associative and self-modifying. Sometimes the associative feature of connectionism makes modification difficult. For example, what experiences might a teacher give the student for whom air is "empty"? Simply defining "empty" and giving verbal examples may not work; it has meant "nothing at all" to the student for nearly two decades. For greater learning, chemistry teachers usually provide experiences about phenomena. Contents of an inverted cup under water remain drv. A released balloon whooshes through air. Gases such-as iodine vapor are wlored. An evacuated vessel, opened under water, fills rapidly Chemistry teachers develop a model for gases that differs from the one students hold. First they deal with the meaning of "empty" (low vs. zero gas density), then with matter whose particles are very far apart. The teacher provides experiences to refine the concept of "empty". Since the students' connectionist systems must be modified, their attention is necessary. Learners do have some control over attention and memory. Confronting learners with a discrepant event, whose outcome seems intuitively obvious but is not. induces self-modification within a connectionist system.' For each phenomena suggested above, challenee the students to oredict an outcome before thev make &servations. This best for bringing abbut longterm changes. Do Many Threads Sew Faster Than One? There is a national thrust to develop a core science curriculum that transcends the boundaries of traditional disciolines where students would take science 1and science 2 in&ead of biology and chemistry. For beginners, a curriculum in which diverse tooics are introduced concurrentlv will be more efiectivc than onc where toplcs arc introduced intensivclv and seauentiallv. We oosit that the more associations already in place, the easier to make new associations. Take advantage of the distributed nature of the system: Extend several networks rather than swamp one network by trying to connect it with many new associations. 750
Journal of Chemical Education
Aeain. the time reauired to make lastine associations is progably hours or days, not weeks. When ekellent instruction. which demands hieh learner outout with much shaoing ikedback, is providA in all cunicha, we do not expe'ct anv multitooic effect to be laree. Adopting a cumculum is deciding what students should know. Neither approach is intrinsically more easy.
-
Naive Theories Revisited A connectionist system is associative. Much appears .. about naive theoriesjn the literiiture, especially concernmg physics. In this research, pcrsnni naive about a particular conceot or content area are interviewed about that material. donsider those students who report that air (a mixture of gases) is "empty" and contains "nothing". Since most humans can lead their lives without ever dealing formally with gases, such a reoresentation mav last a lifetime and require nomodificatibn. After all, ev&yday gas densities are much lower than liquid or solid densities. Relatively speaking, assigning a zero density is an excellent zeroth approximation. Rather than describe these students as learners with naive theories. we sueeest that these are naive learners. learners who lack suitable experience, or learners with no theories. as scientists usuallv use the term. When interviewed about a particular concept (gases, empty), their associative recall mechanisms make their explanation . 123, describes a phen~~menon sound like a r h e o ~Norman that follows funrtioning ol'a connectionist svhtem that he calls deliberate consci&s control. We susp& that interviews with "naive" students involve verbalizations under deliberate conscious control: They follow the parts of their previous experience, knowledge, and learning that teachers wish to change. Our way to deal withlearners who produce naive theories is to provide more content. Of course, the content instruction must demand output and must provide performancerelated feedback.
--
Recalling versus Creating A connectionist system is parallel, self-modifying, content-addressable, and associative. When the content of a test item is perceived so that it is routed to an address which strongly associates it with a response item, the resulting examination performance is very good. With connectionism, when the test reflects well the desired performance outcome, teach closely to the test. (This is exactly how we believe technicians performing drug analyses should be trained.) The human species has flourished because of its creativity. "An idea before its time" expresses the difficulty humans sometimes have with creativity. connectionist systems didn't evolve for creativity; they evolved for recalling. Creativity is probably a special outcome, not the norm of a connectionist system. Connectionism is the key to survival for an organism, providing a means to deal with a slowly changing environment. The fast-changing environment of modern societv is not the one in which the connectionist system evolve&. Creativity is in demand in a society that runs more on chanee than on "business as usual". Problem solving is the call Z t h e day for schools, and teachers try to stimulate creative problem solving. The connectionist system is parallel. Agiven set of inputs can activate many associations. A creative person can use the parallel feature to allow one set of inputs to activate an association that is not usually associated with them. A creative problem-solver can see a relationship between inputs and a unique solution, even when those inputs are embedded among many confusing attributes. For example, a clever synthetic chemist can see the juxtaposition of
atoms in a s i m ~ l model e system and imagine them embedded within a complex organic structure: The Nobel prize often eoes to ~ e o ~who l e have access to only the same data as therest oius,-but they see that data i n a new way that we do not. For instruction, there are a t least two ways to enhance the parallel use of a connectionist system. Fading is teaching concepts cleanly and then embeding them in contexts of increasing complexity (24). Brightening is presenting a complex situation and reducing the complexity until the target concept is recognized. Brightening is part of many discovery-oriented laboratory activities. Fading is sometimes built into series of related lecture demonstrations. One of the authors gives a lecture demonstration series with about 20 different experiments related to density. Robert Becker (25) has devised a series of 15 different Cartesian diver experiments. How will students learn to self-modifyor use their parallel systems unless teachers demand outputs and provide appropriate feedback. While tests present an opportunity for this, there are other possibilities. Group activities, perhaps in laboratory courses, may provide the best environment for nurturing creativity, especially in the sciences. Qualitative and quantitative unknowns present rich possibilities that students really seem to enjoy. Teachers shouldn't expect students after one afternoon of workine with a small set of data to discover what Boyle or ~ e n d e i k e vdid. Were Boyle and Mendeleev slow,-or is discoveryjust easier when someone knows the answer? The teaching of problem-solving skills could be improved. Motivation Some neurons, when they fire, send pain as a message. Organisms act to avoid pain, and this affectsbehavior. The brain produces chemicals that mediate pain (10).There are reward pathways in the brain which, when electrically stimulated by inserted electrodes, produce pleasure. Behavior frequencies are high when pleasure is expected and low when pain is expected. Details of the cognitive learning system are becoming clearer-how one learns sights, sounds, smells, and concepts. The motivational system is chemically and physiologically similar to the cognitive system: They occupy the same cranial space and both have synaptic variations as the onlv mechanism for recall and change, In 20 years this may beconsidered the last time in histoh when the cognitive and motivational svstems were generally thought to be different. Operationally, coupling behavioral response to a reward leads to conditioned learning. Shaping responses in learners remains a major challenge in-teaching. Cognitive instructional strategies are fairly clear and straightforward. Motivational strategies are much less clear. A complete learning model must deal with motivation. Connectionism, as appged to nonhumans and to ancient humans, is a learning system designed to enhance survival. Survival is not really an issue in the classroom just as creativity was not an original feature of connectionist evolution. Perhaps the strategies used to enhance cognition (shaping, brightening, fading) can also enhance motivation. I t is our experience that highly motivated learners generally perform better than unmotivated peers. Successful teaching is inspiring sustained motivation in the learner's behavior. Some teaching strategies and curricula present science in societal contexts. We do not endorse these strategies broadly, but reserve judgement for individual cases. These strategies can provide opportunities for high internal motivation. Children want to mature. Embedding content in everyday contexts provides an especially powerful environ-
ment for students to self-motivate and act out adult roles. Anv dilution of content is offset by an increased rate of learning. When appropriate issues bf other disciplines are incorporated, a win-win situation occurs. Science, mathematics, and engineering change. Computers are expected to change both the way in which science is done and the meaning of being a professional scientist. Innovative STS curricula succeed if they maintain a large flow of committed and respected young persons who are ready for training as professional scientists and engineers. What's Different? A reader should ask "How does learning about connectionism benefit me? Won't my students still have to learn the same thines." ~--We will compare connectionism to two learning paradigms that were prominent during the last two decades. In behaviorism,teachers state explicitobjectives,observe behavior, and then shape it using rewards (and punishments) to meet the objectives. The method works especially well when mar transfer (when teaching and testing are similar to real-world situations) is desired and effective rewards and ~unishmentaexist. "Effective" has a n oDerational definit:on that deals with outcome behavior. In Piagetian strategies (learning cycles), the teacher encourages exploration, possibly with discrepant events, and ~rovidessome backmound. then a ~ ~ l i c a t i oPossibly n. less hnctional at near &ansf& in co&rage, these strategies cause change. ~onnecti&ismrequires that teachers assess student's -maso.of content before deciding whether their goals include near or far transfer. All inst&tion uses feeaback and is designed to achieve high rates of performance. For far transfer, feedback might be less evaluative than for near transfer; open-ended, subjectively graded activities might be undertaken. We expect that future curricula will present highly conceptual material earlier, using instructional strategies not found in New Math or ChemStudy. Research on teaching today is highly input-output oriented. One tinkers withaninput, thenmeasures anoutput. With connectionism, a completely different type of teaching research will emerge that focuses on watching learners as they learn, studying individual physiological responses to variations ininstruction. Some ofthese measurementsmay be made a t CAI terminals. Understanding connectionism is analagous to understanding chemical reaction mechanisms. Once the mechanism is understood, catalysts can be developed. One channel can be favored over another, and so forth. When revising this manuscript, we came upon the following quotation by Dick and Carey (26): ~
~
~
-
A
components... which, when present, almost always will facilitate learning. These three components are those dealing with motivation. orerequisite and subordinate skills, and practice and feedback Especially when near transfer is involved, we think that is excellent advice. Demonstration of Neural Networks Notions of how feature detectors become connected may not seem very credible without a concrete model. For readers interested in pursuing the technical areas described in this article, we have developed a HyperCard-based model of a Kohonen network. This stack is offered with related information about neural networks. The network learns to recognize randomly presented patterns. It maintains a record of its learning activity a t a synaptic level. I t demonstrates that a system with no preknowledge of your input can form appropriate responses to it. The version available at t h ~ writing s rrmemhers pictures of"fnces", 64-
Volume 68 Number 9 September 1991
751
x 64-pixel displays. After a delay, i t will recall those faces. Then, given part of the face, it will recall the entire face. Well send a copy to you with our faces, and you can add faces of your own and see how it works. We offer this stack to illustrate that one can model neuronal .behavior in a straightforward manner usina comuuters. If vou k e e ~ the problem small enough, the reiults i r e drama&! The promam requires a MacintoshcomDuter with a hard drive-and ~ ~ ~ e k version a r d 1.2.2.-To obtain t h i s Hypercard stack, send a check for $5.00, made out to the University of Nebraska, to D. W. Brooks, Science Center, 118 Henzlik Hall, UNL, Lincoln, NE 68588-0355.
W.H. Freeman:NewYork, 7. N*"ta,W. J.H.;F~*ag,M.Fundom"taiN~umauiimy: 1986. 8. Nadel, L.: Cooper, L. A.; Culicover, P.; Hsmish, R. M., Ed-a:Neuml Connections, Menial Computation: MIT Ress: Cambridge, 1989. 9. M m , D. Vision W. H. Reeman: New York, 1982. ondBehouior:W H R a e m a n : 10. Blaom,F.E.;lazerson,A.:Hofstadter,L.Bmin,Mind, NewYork. 1986. 11. Dennett D. C. T i m e and the Obeerver'. Cedrie Evans Lechlre. Deoalimmt o f
. . .
13. Bouchard, T. J;Lykken, D. T.; Mffiue, M.: Segal. N. L ; Tel1egren.A. Sclenee 1990, 250.223. 14. Criek,F. H. C. S c i s n t i f ~ A m r i m n 1979,241,3,219-232. . 15. Caudill. M.; Butler. C.Ndumllylnfelllgent Sys&ms; MIT Press: Cambridge. 1989. 16. HaU. S. S. S c G m 85. 1985.5.1099 ~17. ~ c ~ l e l l a nJ. d ,L.; Rumelhart, D. E.; Hinton, G. E, in referenee 5, pp 3 4 4 . Literature Cited Is. Kohonen, T. Content-cuLirrssobY memory: Springer-Verlag New York. 1988. 1. Edelman, G. M.NeurdDominism. Tb Thmry ofNeurnu1 Gmup Sekflon;Basic 19. Nowak,L.:Bregestouslu,P.:A~chhhhP.:Herbet,A.;Pmhiantl.A.Nhttf,1984,307, Bwks: New York, 1987; p I . 462465. 2. B m e . J. H.:Berru, W. ONovml ModPlsofPhfrcity: Ezp"mento1 Md Tksomtien1 Raichle,M.E.Sebm 1988,%0,1627-1631. 20. P0Sner.M. I.;Petersen,S.E.:Fox,P.T.; 21. Wollman, W.T.;Chen, B. SeiomEducotion. 1982,66,717-730. 22. Lawsan, A. E. Presented st the National Science Foundation Roje"t Ere