What is this? Can Connectionist Does the is disputed Can Computers Think in Images? 5 Networks Think? by Can connectionist networks subsymbolic account exhibit systematicity? offer a valid account 69 Computers can't understand images. Computers can't think because they can't use images in the way that people do. Computers can only deal with formal symbolic information. ? 1998 Can computers The History and Status of the Debate — Map 5 of 7 33 Tim van Gelder, 1990 func • tion • al com • po • si • tion • al • i • ty: of connectionism? Note: "Image" in these arguments usually refers to an "eyes-closed" image, which is imagined in one's mind without the real object necessarily being present. Sometimes the term is also used to describe is supported by recognize Gestalts? Functionally compositional representations avoid the connectionist A representational scheme that can produce complex an "eyes-open" perception. This info-mural is one of seven dilemma. Smolensky and others have developed "functionally compositional" representations from parts and that can decompose a representational schemes, which can account for systematicity without complex representation back down into those parts. An Issue Map™ Publication 32 implementing a classical architecture. Such representations can be built up from parts and can be broken back down, but they are non-concatenative— that is, they don't explicitly contain or "token" their parts. The process can be repeated to create increasingly complex representations. 57 Paul Smolensky, 1988b Can images be realistically The subsymbolic paradigm. The fundamental level of analysis Connectionist Supported by “argumentation maps” in a is supported by for studying the mind is the subconceptual level, which describes 83 Edwin Boring, 1946 representations avoid the "Tensor Product Representations Avoid the Dilemma," Box 38. fine-grained subsymbolic activity in a connectionist network. Classical is Gestalt theories are consistent with computer inference. dilemma. Connectionist disputed 82 symbol manipulations are rough approximations of subsymbolic From a computational point of view, perception involves unconscious Start Here representations can exhibit is supported by Note: Van Gelder points out that Gödel numbers also exhibit nonclassical con • cat • en • at • ive com • po • si • tion • al • i • ty: activity. Subsymbolic activity is, in turn, an abstraction from neural activity. Explanations of all 3 kinds of activity— symbolic, by Gestalt recognition is impossible for computers. is inferences similar to those made by electronic computers. From a represented in computer arrays? systematicity and related constituent structure. See sidebar, "The Steps of Gödel's Proof," on Map 7. disputed Gestalt point of view, perception results from the operation of Gestalt recognition involves the subsymbolic, and neural—are legitimate for cognitive theory, so long phenomena without Functional compositionality, with the added feature that complex by dynamical fields of force in the brain. Both theories provide legitimate series that explores Turing’s implementing a classical representations explicitly contain their parts. Parts are "literally immediate comprehension of a present" or "tokened" in complex representations. Classical as the proper importance of each is understood. pattern as a unified whole. scientific explanations of the phenomenon of perception, just from symbolic architecture. Note: Also, see sidebar, "Postulates of the Subsymbolic Paradigm," different points of view. There is no fundamental contradiction symbolic representations are concatenative in this sense. Computers can only sequentially Unmapped Territory 34 Keith Butler, 1991 on this map. Paul Smolensky 71 Michael Tye, 1991 process the components of a between the 2 approaches. Unstructured representations can account for Images are interpreted 72 Anticipated by Michael Tye, 1991 73 Michael Tye, 1991 pattern; they can't recognize Additional 70 symbol-filled arrays. Connectionist objection to Tye's image theory can Gestalt wholes. systematicity. Semantic structure can be represented question: “Can computers think systematicity Tye's theory. Tye's theory is 1 Alan Turing, 1950 I believe that at the end of the century ... one arguments is supported by by the causal history of an activation pattern. For example, "John loves Mary" is an unstructured activation pattern that can be traced back to earlier causal effects of "John," Role: Subject Tensor Product Representations In a tensor product representation, vectors representing roles (e.g., "subject" and "verb") are combined with vectors representing role fillers (e.g., "John" 58 George Lakoff, 1988 Subsymbolic representations and the Images can be represented in computers by filled cells in an array. is supported by Images are two-and-one-half- dimensional arrays of symbols that encode two- is disputed by flawed by its classical conception of memory as storage. Memory is is disputed by accommodate connectionist memory. The theory does not require Note: Map 3 contains a variety of Gestalt-inspired objections to symbol systems. is disputed 84 Douglas Hofstadter, 1978, as articulated by Joseph Rychlak, 1991 Gestalt properties of images can be implemented recursively. Recursive loops can explain the Gestalt figure–ground Yes, machines can will be able to speak of "loves," and "Mary" activation patterns, and perhaps even further back to microfeatures of those. By virtue of causal and "Mary") by taking their tensor product. A tensor product is the vector that results from multiplying each element of one vector by each element sensorimotor system. The subsymbolic paradigm, supplemented by an account of the Cells in a matrix inside a machine's memory function as dimensional information about physical surfaces along not stored in a classical representational manner, but is that images be stored representationally. The by relationship. Recursive loops allow the background of an image to be constructed as a complement (or opposition) to the figure. and/or will they ever be able history, unstructured connectionist representations can body's role in meaningful cognition, overcomes if they were arranged in a with viewer-centered reactivated in a connectionist discrete parcels of information machines thinking 2 1 4 of the other. In a connectionist network, this can be implemented by feeding (or will be able to) without expecting to be account for systematicity and related phenomena. So, Fodor and Pylyshyn are right to claim that connectionist representations are unstructured, but wrong to claim that 1 x x x 2 x the role and filler vectors into separate input layers that connect at a set of multiplicative junctions. is supported by problems that plague classical AI. In classical symbolic AI, representations are only meaningful by virtue of arbitrary associations with things in visual array. Images correspond to filled cells within such an array. Array- information, such as depth and relative orientation of surface features. Such arrays dispositional fashion, based on the strength of connections between processing nodes. contained in images may be stored in a connectionist dispositional system. is disputed think. A computational contradicted. such representations can't account for systematicity. 2 1 4 2 Once a filler has been combined with a role it may be combined with other the world. Connectionist representations (i.e., based sensory patterns are have an attached sentential by to?” Argumentation mapping is activation patterns), by contrast, are intrinsically interpreted by higher-level interpretation that specifies is supported by 1 x x x x role or filler products by vector addition (where corresponding elements are simply added together). meaningful by virtue of nonarbitrary connections systems. their representational content. is 74 Michael Tye, 1991 35 Paul Smolensky, 1988a disputed Connectionism cannot accommodate all system can possess all is disputed by is supported by The coffee story. To see how distributed representations can encode compositional structure, consider a connectionist representation Filler: John 0 x 2 x 1 x 4 x 2 2 142 2 142 1 0 1 1 2 0 0 1 00 1 1 01 1 4 31 64 42 57 to eyes, ears, limbs, and so forth, which in turn are dependent on the surrounding world. Note: Also, see sidebar, "Postulates of Experiential Note: This argument is sometimes referred to as the cathode ray tube, or CRT, is supported by by the evidence about images. Evidence suggests that image generation is a of coffee. The representation can be obtained by subtracting a vector 0 0 0 0 0 000 + 1 1 0 2 + 01 3 0 = 12 32 constructive process that draws on separate packets 85 Joseph Rychlak, 1991 important elements of Realism, on Map 3. metaphor. Arrays are also a method that provides: human thinking or Key: Icons for Conceptions of Thought The following icons are used throughout this map to symbolize different conceptions of thinking: 38 Paul Smolensky, 1988a Tensor product representations representing the microfeatures of cup from a vector representing the microfeatures of cup-with-coffee. The resulting representation of coffee is context sensitive in a nonclassical way. It is a representation of coffee in the context of a cup. 1 x 2 x 1 x 4 x 2 2 142 John (subject) 0 1 4 0 Mary (object) 11 0 1 Loves (verb) 33 83 John loves Mary 59 Douglas Hofstadter, 1988 Common sense and connectionism. Commonsense reasoning in humans is best called matrices or surface matrices. of information stored in long-term memory. The connectionist dispositional theory lacks an account of how those packets can be stored discretely. is supported by The recursive interpretation of Gestalt images gives the wrong picture. Recursive loops do not allow the background of a figure-ground structure to have independent properties. But a principle property of Gestalt images is that their backgrounds are symbols in their own right. This shows that Gestalts are not recursively constructed, as Hofstadter claims. - a method for portraying major Note: Also, see sidebar, "Postulates of the Dialectical Paradigm," on Map 3. understanding. Symbolic accounts of the mind avoid the dilemma. Tensor product representations (see sidebar, "Tensor Product Representations," on this Cup 36 Jerry Fodor and Brian McLaughlin, 1990 requires is supported by explained by a "mental topology" of conceptual "halos." For example, the concept of "contact" is surrounded by a halo of 75 David Marr and Keith Nishihara, 1982 Visual perception utilizes two-and-one-half-dimensional arrays of symbols. Visual processing occurs though stages of construction. 86 Rudolph Arnheim, Connectionism map) have constituent structure but The regress of contexts. requires requires concepts that include "call on the phone," "go 1. Visual information is gathered by recognition of intensity differences. Postulates of Gestalt Psychology with Cup Coffee is Smolensky's coffee representation leads see," and "write." A mental topology can be 2. A two-dimensional "primal sketch" is made, which orders the visual information into a representation of edges and other 1969 are distributed and context sensitive coffee disputed Computers can't philosophical, political, and Alan Turing Neuroscience in a nonclassical way. Vectors to an infinite regress of representations. explained in the subsymbolic framework, surface details. by Coffee depends on a higher-order where conceptual halos are understood as 3. From that primal sketch a two-and-one-half-dimensional sketch is constructed to include depth and movement features the process images from 1. The whole is different from the sum of its parts. representing roles (e.g., subject or the top down. Humans object) and fillers (e.g., John, Mary, Microfeatures representation of cup-with-coffee. But Cup overlapping regions in an abstract space. physical surfaces of the object. loves, etc.) can be bound together by upright container then cup-with-coffee presumably with Cup of Note: Also, see the "Can a symbolic 4. Finally, the two-and-one-half-dimensional sketch is compared with other stored two-and-one-half-dimensional sketches. recognize images from 2. The whole is studied in terms of its form or organization. A Gestalt—which in German means "pattern" or "shape"—is best understood as kind of perceptual configuration. 1 1 0 Coffee coffee on the top down, taking a network that takes the "tensor minus equals depends on some further representation, coffee knowledge base represent human The result is a three-dimensional structural description with a hierarchical structure. hot liquid 1 0 1 the table Douglas Hofstadter them in as a whole and product" of the two vectors. The such as cup-of-coffee-on-the-table. And understanding?" arguments on Map 3. pragmatic debates vectors that result are then added finger-sized handle 1 1 0 so on. then focusing on details. 3. Laws of grouping describe the general patterns of organization obeyed by perceptual wholes. together to produce complex Computers only burnt odor 1 0 1 recognize images from 60 Walter Schneider, 1988 Are images is supported by representations. No grouping. 40 Terence Horgan and John Tienson, 1992 Specially crafted networks produce symbolic the bottom up, processing processing. Modeling symbolic activity requires History of the Image Debate local features individually Distributed parts Nonclassical constituents can be causally effective. The constituents of a Proximity: Elements that are closer to one another are 39 Jerry Fodor and Brian McLaughlin, and then aggregating - a summary of an ongoing, tensor product representation may have causal powers in the same way that naturally connectionist architectures that are "hand-crafted to produce For most of the 20th century, imagery research enjoyed little grouped together. 1990 of tensors lack favor, due to the dominance of behaviorist psychology. John them together. causal powers occurring (nonclassical) constituents do. For example, the wake left by a motorboat is symbolic-like processing" (p. 51). Symbolic processes quasi-pictorial? Either Or is Constituents of tensor product is a complex superposition of wave motions, none of which are explicitly contained don't "emerge" from connectionist networks (as Smolensky Watson (1913) led the behaviorists in rejecting imagery research, representations lack causal disputed Similarity: Elements that are more similar to one disputed 3 164 or tokened on the surface of water. Nevertheless, nonclassical constituent waves have claims), they must be explicitly built in. In the future, because it involves nonobservable, introspective evidence. another are grouped together. is 31 Jerry Fodor and by powers. The parts of a tensor product by is Zenon Pylyshyn, Connectionism can't account Connectionism accounts for 4 257 numerous causal effects, including "setting a buoy to bobbing in a certain way, knocking connectionism will have to move beyond simple disputed for systematicity and related representation are not explicitly During the 1960s, through the efforts of Allan Paivio and disputed feedforward networks to accommodate richer brain-inspired major philosophical debate of contained or "tokened" within it—the 2 by 1988; systematicity and related 1 232 down a skier, and contributing to the destruction of a sandcastle" (p. 212). Closure: Elements that form closed units are grouped phenomena (productivity, phenomena by using structured architectures. Connectionism, as part of a "team of concepts others, image research was revitalized. Images were given a by Connectionist networks can think. Connectionist networks can possess all important elements of human thinking or Jerry Fodor and parts have no independent status in the 3 383 is supported by together. compositionality, and representations, is and tools," should ultimately model such phenomena as Imagine 2 triangles. Take one, role in explanations of learning, memory, perception, and other understanding. Connectionist networks are characterized by: History of Connectionism Brian McLaughlin, complex representation. As such, tensor 77 Geoffrey Hinton, • an ability to learn via training, rather than explicit programming inferential coherence), disputed Either.... single-trial learning, attention, multispeed learning rates, flip it over, and superimpose psychological processes. Good continuation: Elements forming continuous lines 1990 product constituents lack individual Explicitly tokened 1979a • parallel and distributed processing Several historical precursors to connectionism are cited in the literature. It is sometimes The connectionist in which case causal powers. Classical constituents, parts of tensors by and working memory. is The quasi- is supported by it on the first. This is what or curves are grouped together. • neural realism, or at least neural inspiration claimed that Aristotle, with his focus on learning and intuition, was a proto-connectionist. in which case disputed you see. In the early 1970s, a new cognitivist critique of imagery was the 20th century dilemma. The by contrast, have causal powers by virtue have causal powers pictorial view • fluid tolerance of noisy or incomplete data The works of the British Empiricists (John Locke, George Berkeley, and David Hume) connectionist approach connectionism is a mere of being explicitly contained in complex by can't explain launched by the advocates of cognitive psychology, notably Common region: Elements that are located in the same • superior performance on perceptual and motor tasks are also viewed as precursors to connectionism, given that associative links between to cognitive science is connectionism is inadequate implementation of the classical 61 Marten den Uyl, 1988 some image Zenon Pylyshyn. The critique emphasized the importance of 87 Stephen Kosslyn, 1994 perceived region tend to be grouped together. Note: This general characterization of connectionism is intended to highlight those aspects of the field that are relevant to this ideas are similar to weighted links between nodes. (Note that the Empiricists are also as a theory of cognition. architecture. representations. John Loves Sally The subsymbolic paradigm needs to symbolic structures, or propositions, in the explanation of Top-down processing impaled on the horns of effects. Some 78 Geoffrey Hinton, image effects. A psychological debate ensued that is still of images has been a dilemma—it is either map. Few connectionists would actually claim that "connectionist networks can think," because connectionist networks are pinpointed as precursors to classical AI; see sidebar, "History of the Symbolic Data is analyze between-module structures. Connectedness: Elements forming a uniform disputed 76 Stephen Kosslyn effects of 1979b active. implemented in usually regarded as simulations of neural networks in the brain, which is where the real thinking is understood to take place. Assumption," on Map 3). inadequate as a theory is Smolensky's analyses apply to single processing manipulating images Structural connected region are grouped together. - a new way of doing intellec- by modules in the cognitive system, for example, and James Pomerantz, 1977 Imagine 2 parallelograms that computer programs. of mind, or else it is a disputed Images are quasi-pictorial cannot be explained descriptions The current imagery debate is primarily a debate in psychology Top-down processing of Activity of a Formal analysis of neural networks was pioneered by biologists and cyberneticists in the mere implementation of In Either Case by to a "smell module" or a "depth-perception by interpreting them are drawn together and affect imaging Thinking = connectionist network = = 1940s and 1950s. Warren McCulloch and Warren Pitts (1943) developed a "logical calculus" of neural activity, which showed that neuron-like elements could compute the classical module." However, to analyze the simultaneous operation of multiple modules will require new representations. Images are quasi-pictorial entities with spatial as picture-like, tasks. This superimposed. This is what about the nature of mental images, how they are processed, and how they are implemented in the brain. Some of the debate imagery has been formally described using a theory of 4. Frames of reference determine how a set of stimuli will be grouped into a perceptual whole. architecture. "The mind cannot be, in its general structure, a connectionist symbol-filled arrays. you see. logical functions (see "The Logical Calculus of Neural Activity," Map 3, Box 6). 37 Noel Sharkey and Stuart Jackson, mathematics beyond what Smolensky describes. properties that correspond to those example shows that is represented on this map in the "Is image psychology a valid processing subsystems. 5. A Gestalt pattern can be perceived either as an independent object (figure), network" (Fodor and Pylyshyn 1988, p. 33), and the classical 3 elements of a of their correlated physical objects. However, if we how a shape is approach to mental processing?" arguments. This theory is tual history. 1994 1 element of a context- interpret images as or as the surface or background behind the object (ground). Certain laws What does it Donald Hebb (1949) proposed that learning in neural networks takes place when 2 symbolic paradigm is still the best explanation of mind. Images possess structure by virtue perceived depends mathematically precise determine whether a region will be seen as a figure or ground. Weight representations avoid the context- independent is structural connected units are simultaneously active. Frank Rosenblatt (1959) may have been the of their ties to higher perceptual on the structural For further discussion of the history of imagery in psychology, enough to be simulated in Postulates of Connectionism John loves Sally regress. One part of a context- dependent weight disputed 62 Richard Golden, 1988 processes. descriptions, those description that is see Allan Paivio (1971, pp. 2–8) and Stephen Kosslyn (1994, computer programs that first modern connectionist. His perceptrons (one-layer feedforward networks) could 41 David Chalmers, 1990b dependent vector representation is a activation representation by is effects can be easily 6. According to the principle of Prägnanz, or closure, learn to recognize patterns reasonably well, and he developed early forms of several Implicitly structured .2 .1 .7 1 disputed Statistical rationality needed. Smolensky needs to describe the subsymbolic Supported by used to construct it. However, the shape is the pp. 1-4). recognize patterns by a ambiguous stimuli are interpreted in the most simple, 1. There is a set of elements called units or nodes. context-independent weight representation paradigm more precisely by stressing the role of statistical inference in explained. Such modern learning schemes. Rosenblatt also introduced the use of computer simulation representations can engage in representation. Weight representations by "Image Psychology," Box 88. same in both cases. process of top-down regular, and symmetric pattern possible, based on Sample Node 44 Jerry Fodor and Zenon Pylyshyn, connectionist systems. To do this, Smolensky needs to recognize that: structural hypothesis testing. into neural network theory. is supported by structure-sensitive processing. A available information. • "logical inference is a special case of statistical inference;" 2. Nodes pass signals (numerical values) to each 1988 "provide a contextually stable .6 descriptions network can be trained to transform • "rational connectionist models are statistical inference mechanisms;" other. Connectionism is representation of constituents from represent parts and Postulates of Quasi-Pictorial Activation value From the late 1950s to the 1970s there was a lull in connectionist research, which resulted Unmapped Territory distributed representations of passive • "continuity is necessary for representing partial (real-valued) beliefs" (p. 35). associationism. Processing units which context-dependent expressions objects as nodes and 3. Nodes are connected into networks, which Weights from the early success of AI combined with Marvin Minsky and Seymour Papert's critique in a connectionist network are sentences into the active voice, and vice may be constructed" (p. 168). spatial relations as Image Psychology 7. Psychological systems that reduce thinking to discrete operations on data, such as structuralism, Output of perceptrons, which was developed throughout the 1960s (but not published until 1969). versa (thereby exhibiting systematicity). behaviorism, and AI, are inadequate, because they fail to account for Gestalt properties of perception. resemble neural networks in the brain (see function causally connected by associative Additional labeled arcs. 1. Images, like pictures, have spatial properties. Images sidebar, "Connectionism And The Brain," on Input Their critique showed that one-layer networks couldn't compute certain kinds of functions links. But associationist theories associationism This ability shows that the implicit is and pictures may not be fully spatial in the sense that .3 (see "One-Layer Perceptrons Can't Compute Certain Functions," Box 10). are "cognitively weak"; they cannot structure of its representations can be disputed Note: Presented here is primarily the Gestalt theory of perception. Applications of Gestalt theory this map). arguments .5 -1 0 1 physical objects are, but they still have some degree of to learning, motivation, education, and social psychology have been excluded. 0 Output account for systematicity and related casually effective in processing. by spatiality. Images are quasi-pictorial entities. is 79 Anticipated by Stephen Kosslyn and James Pomerantz, 1977 contain? .8 0 .6 .9 Connectionism began to resurface in the 1970s. One of the classic papers of the new phenomena. Symbolic AI, however, Sally is loved by John 4. When a set of signals reaches a node, they are can explain systematicity with disputed Quasi-pictorial images face an infinite regress. Pictures are intrinsically perspectival, Authors on this map whose work draws on Gestalt principles include Hubert and Stuart Dreyfus, multiplied by weights and then are added 1 generation was Jerome Feldman and Dana Ballard's "Connectionist Models and their is supported by 2. Images are analogous to perceptual presentations and .6 Properties" (1982), which introduced the term connectionism, outlined some of the basic reference to structured by and so the notion of a quasi-pictorial image makes the assumption of a point of view, or Edwin Boring, and Rudolf Arnheim. Other notable Gestaltists include Kurt Koffka, Wolfgang Köhler, together to produce an activation value. representations and structure a "mind's eye." Because of that assumption, the use of quasi-pictorial images leads to an are processed (at least in part) by the same systems. Kurt Lewin, Max Wertheimer, and Edgar Rubin. properties of connectionist models, and argued for their superiority over classical AI systems. The major publication of this new generation was David Rumelhart, James sensitive thought processes. 63 Paul Smolensky, 1988b infinite regress of interpretations. The image must be interpreted by the mind's eye, but 5. The activation value is then passed through an Sample Output Functions Constituent Structure of Mental Representations then that perspective must also be interpreted, because the mind's eye then becomes part 3. The cathode ray tube (CRT) metaphor provides a Stephen Kosslyn Drawings adapted from Irvin Rock and Stephen Palmer (1990). output function, which decides what final value McClelland, and the PDP Research Group's Parallel Distributed Processing volumes Too early to throw out computational approaches. Some processes, such reasonable initial model for properties of mental imagery. (1986), which sold out upon publication and which rallied many AI researchers to the Classical symbolic theories as motor control, have yet to be modeled in a statistical framework. So even if of the image, and so on ad infinitum. Therefore, the idea of picture-like representations a unit should output. Threshold function Sigmoid function postulate a language of thought in the mind is incoherent. The metaphor may have to be abandoned at some point, connectionist camp. The writers described modern connectionist research in detail and is supported by Golden is right about the centrality of statistical inference, it is too early and too but it contains many essential traits of imagery (see argued for the superiority of connectionism over classical symbolic AI. 45 Brian McLaughlin, 1993a (see "The Language of Thought," Map restrictive to exclude other computational approaches from consideration. 6. Vectors (sets of numeric values) are the basic 3, Box 68), according to which Image in Mind's eye #2 Mind's eye #3 "Images Can Be Represented In Computers By Filled representational medium of a connectionist The burden of proof is on Implemented Model Cells In An Array," Box 70). Renewed interest in connectionism was answered by the classical AI camp with a series connectionism. The burden of proof complex mental representations 42 Brian McLaughlin, 1993a is mind's seeing mind's seeing mind's Image Are images less network. Vectors enter the network as input, If activation is above some A continuous version of the are built up out of more simple 43 John Pollack, 1990 Chalmers's representations lack disputed eye #1 eye #1 eye #2 seeing are processed through vectors (and matrices) of polemics, in particular a special issue of Cognition (1988), which contained 3 long is on the connectionists to explain complex representation 4. Images play a functional role in cognition. They are level, fire one value, threshold function, which representations. Complex The RAAM Network. Passive–active transformations can be achieved, syntactic structure. The representations 64 John McCarthy, 1988 by mind's eye #1 108 Zenon Pylyshyn, 1973 of weights and connections, and new vectors critiques of connectionism by prominent cognitive scientists. The gist of the 3 Cognition systematicity and related phenomena ("molecular representation") not epiphenomenal. Dual codes are too otherwise fire another value. outputs one of a range of arguments is the same: connectionism is a revival of associationist theories, which are without implementing a classical representations can themselves be in part, via a recursive auto-associative memory (RAAM) network. A RAAM used by Chalmers lack syntactic structure, even Connectionist networks can't quickly elaborate their capacities. are produced as a result. combined to form higher-order is Whereas humans (and AI systems) can quickly elaborate their capacities, indeterminate to encode values rather than just one of cognitively weak; such theories can't account for the kind of systematicity and generativity architecture. John Loves Sally network compresses constituent vectors into complex representations, and if they represent sentences that possess representations. decompresses those complex representations back into their constituent parts. syntactic structure. Because they lack real disputed connectionist networks can't. For example, when an English-speaking human is is 5. Images are constructed from "chunks" that are stored knowledge. Dual codes are too two values. fundamental than 7. Neural networks are trained to compute vector that symbolic accounts explain so well; and, if connectionism can account for such by disputed disputed separately in memory. Images are not pulled from ambiguous to provide for phenomena, it offers a "mere implementation" of the classical view (see "The Past-Tense The processes of compression and decompression can be repeated to deal syntactic structure, such representations cannot is given a simple rule for Chinese pronunciation (e.g., say "ch" when you see to-vector functions. That is, they learn to When a complex representation with arbitrarily complex connectionist representations. participate in structure-sensitive (or "Q"), he or she can immediately use the rule to speak differently. This kind by by memory as fully formed units. correspondences between pictures convert specific input vectors to specific output Model Does Not Argue Against Rule-Based Explanation," Box 24, and "The Connectionist Connectionist Network Computing Dilemma," Box 31). explicitly contains its parts, those parts syntax-sensitive) processes. of learning could not occur in a connectionist network, which would have to and words. For example, someone vectors. Vector-to-Vector Function 46 Daniel Dennett, 1991b are said to be tokened in the complex John Loves Mary instantly adjust thousands of connection weights. Humans and AI systems 6. Conceptual information in long-term memory can shown a picture of a rose doesn't Cognition isn't always systematic. .2 .1 .7 1 = influence image construction. propositions? representation. The parts of a complex John Loves Sally are "elaboration tolerant"; they can quickly extend their abilities to take into know whether the word to associate 8. Learning takes place at the weights, which are The debate between connectionists and classicists, which has been called a "holy war" is Not all organisms exhibit systematicity of .2 .1 .7 1 0 -1 0 .7 .3 .4 .1 0 representation are usually referred to John Loves account new phenomena. with it is "rose," "flower," "plant," adjusted using a variety of learning procedures, and a "battle to win souls," continues. Although the debate has often been heated, current disputed thought processes. For example, there are Sally 7. Images as reconstructed from perception and memory .3 .2 as constituents. or another word. Moving from generally involving exposure to a corpus of discussions lean increasingly towards "ecumenical" views, which try to give credit to the by creatures that can in some sense think "The Arguments 80 Stephen Kosslyn and James Pomerantz, 1977 81 Stephen Kosslyn, Steven Pinker, George E. Smith, and are intrinsically object-directed, or "intentional." pictures to words and vice versa Input Output virtues of both classical and connectionist approaches (see sidebar, "Spectrum of Positions," constituents of the representation Unmapped Territory sample inputs (see sidebar, "Learning in lion is going to eat me" but cannot think "I about levels are is The mind's eye should be seen as a classification Steven P. Schwartz, 1979 Photographs are not intentional. requires an underlying intermediary Classical theories claim that mental (at the lowest level, "atomic Altogether the seven maps: vector vector on this map). am going to eat the lion." widespread in scheme. The mind's eye need not be interpreted as a location Machine implementations using images have displayed Connectionist Networks," on this map). 1 processes of inference, transformation, representations") disputed code, or interlingua, to mediate the .3 .2 the philosophy Additional by in space. The mind's eye is better thought of as a processor real computational power. Quasi-pictorial imagery does not Note: This set of postulates is also called the "pictorialist two realms. So, dual code theories 1 0 Further history is contained in Rumelhart and McClelland (1986b, pp. 41–44), and in composition, and so forth are structure levels (or "visual buffer") that interprets sensory information in a lead to an infinite regress about a mind's eye. In fact, computer Proponents include David Rumelhart, Paul 1 0 sensitive, which means they operate of connection- view" or the "picture theory." 107 John Anderson, are inadequate. Smolensky, James McClelland, Geoff Russell and Norvig (1995, pp. 594–596). ism and AI, and arguments series of stages. In the process of interpretation, sensory models based on a visual buffer have been highly successful at a 0 1 directly on the constituent structure variety of tasks, even to the point of being able to solve problems 1978 Hinton, Jerry Feldman, Paul and Patricia of representations. go back at least information is classified in terms of conceptual categories Proponents include Stephen Kosslyn, Michael Tye, and A dual code interlingua -1 1 .7 - summarize over 800 major Churchland, Terence Horgan, John Tienson, and 0 to the work of that are correlated with objects and their properties. they were not originally programmed for. Such success could not Mark Rollins. is .7 .5 -1 0 1 result from an incoherent theory. is theory explains David Touretsky. Other notable proponents include David Marr. knowledge as disputed Other connectionist disputed rose?... Jeff Ellman, Stephen Grossberg, and Terrence John Loves Mary 65 Paul Smolensky, 1988b by The conscious rule-interpreter can elaborate its capacities. The one-trial learning process by well as flower?... Sejnowski. -1 .7 47 Robert Matthews, 1994 Other ways of referring to constituent structure include: propositions do. example of Three-concept monte. By the way they pose their • compositional structure • syntactic structure that McCarthy describes is carried out by a "conscious rule-interpreter," which learns new rules and A dual code a species? challenge to connectionism, Fodor and Pylyshyn (and • compositional semantics • combinatorial syntax and semantics can apply them immediately, although at a comparatively slow rate. Eventually, the new rule is moves in the debates threaded Spectrum of Positions consisting of images arguments • language of thought Is image psychology a valid encoded in the weights of the "intuitive processor," which is comparatively fast. How the conscious later, McLaughlin) make it hard for connectionists to • compositionality rule-interpreter would change its weights for one-trial learning is a subject of current research. is supported by and associated Learning in Connectionist Networks provide an acceptable response. While distracting us 106 Zenon verbal strings with talk of systematicity and implementation, they Pylyshyn, can be used to is Connectionist networks learn by incrementally conceal the crucial issue of explanation. It turns out that Implementationalism 1973 encode knowledge. disputed adjusting their weights in response to a corpus the only explanations classicists are willing to accept are 66 Walter Freeman, 1988 Images are approach to mental processing? Such a dual code by into claims, rebuttals, and of sample inputs. The networks are repeatedly fed these inputs until they have been trained to perform as desired. Training takes place via a learning algorithm. 9 Hubert Dreyfus, 1992 Connectionist computers lack a commonsense background. Connectionist networks are unable to make generalizations and classifications in the way human beings do, because they lack our commonsense understanding is disputed by classical explanations. Explanation Systematicity Systematicity and Related Phenomena The ability to think certain thoughts (or say certain things) is If someone can think John Loves Sally Implementationalism The mind, in its general structure, is a symbol processor. Connectionism This is what matters. Forget the brains, focus on symbols! is disputed is disputed by Too much representation, not enough dynamics. Connectionists like Smolensky are right to emphasize dynamics and complex activity patterns. However, Smolensky's notion of subsymbolic representation carries connotations of registration, storage, retrieval, and backpropagation, which do not appear to play a role in the complex neural dynamics of animals, like rabbits. Furthermore, 89 Zenon Pylyshyn, 1973 90 Stephen Kosslyn, 91 Allan Paivio, 1979 is disputed by secondary to propositions. Knowledge is encoded in an unconscious is by itself adequate; it isn't necessary to postulate a more basic level of propositions. counterrebuttals Learning is a central emphasis in of the world. For example, one of the army's early connectionist networks was intrinsically connected to the is only useful as a theory by connectionists rely too heavily on the simple dynamics of equilibrium attractors. The definition of image is too vague. The notion Steven Pinker, propositional 109 John Anderson, 1978 trained to distinguish photos of tanks from photos of empty landscapes. The Computational medium that lies connectionism, and there is a vast technical ability to think certain other she or he should also be able to think of how those symbolic In doing so, they ignore the more complex dynamics of limit cycles and chaos, 88 of "image" has no clear meaning except by association George E. Smith, and theories of imagery Propositional codes are not necessary for literature on the various connectionist learning results were initially promising, but it was later discovered that the network thoughts (or say other things). These are mere which play important roles in the neural dynamics of animals. with the commonsense notion of a picture. But the beneath both translation between verbal and visual codes. Image psychology. made its generalizations inappropriately—by detecting whether or not there Unmapped Territory processes are implementation details. Steven P. Schwartz, give inadequate algorithms. The most widespread learning Systematicity can also be Sally Loves John implemented in the Note: For elaboration on these points, see sidebar, "Postulates of the Dynamical Members of this school notion of a picture is misleading for the study of mind, 1979 language and If it were necessary to have a third, propositional code were clouds in the picture (the pictures with and without tanks had been taken account of basics. imagery. procedure is the backpropagation algorithm, characterized as a kind of mental brain. Approach to Cognition," on this map). of psychology claim that is because it implies that a spatial geometric figure is is Computer is to translate between visual and verbal code, then it on different days). somehow actually present in the brain when we perceive Kosslyn's Imagery by itself would also be necessary to have a fourth code to translate which changes the weights of a feed-forward Additional symmetry. disputed disputed - 97-130 arguments and rebuttals Systematicity Proponents: Zenon images play an essential simulations show disputed computational theory network based on the error generated by a learning is Note: For similar arguments applied to the symbol systems architecture, see the Implementation Pylyshyn, Jerry role in thinking. by images. by how the image by is not of interest from verbal to propositional code, and a fifth to translate "Can a symbolic knowledge base represent human understanding?" arguments A related point is that systematically related thoughts are not related to certain fails to explain either given input. Other important learning arguments disputed Fodor,and Brian 67 Louise Antony and Joseph Levine, 1988; William Bechtel, 1988; Note: Image That definition is too theory can be made to cognitive from verbal code to the new intermediary code. This on Map 3. other thoughts. For example, "John loves Sally" is systematically related to vague. the developmental algorithms include competitive learning and the by "Sally loves John" but not to "2 + 2 = 4". This property of cognitive systems is McLaughlin. B. Chadrasekaran, Ashok Goel, and Dean Allemang, 1988; Carol Cleland, conceptual level psychologists are not precise. Images can science because leads to an infinite regress. origins of images or it can't explain Boltzmann machine learning algorithm. so closely linked to systematicity that it is often overlooked. It has been called 1988; Stephen José Hanson, 1988; Dan Lloyd, 1988; Chris Mortenson, directly concerned with be precisely described as their stimulus 1988; Gardner Quarnton, 1988; Georges Rey, 1988; Jay Rueckl, 1988; subconceptual level the issue of whether computational data human compositionality, but this term is confusing, because compositionality is normally per map conditions. The theory knowledge. is supported by used to refer to the compositional structure of mental states (see sidebar, "Constituent Walter Schneider, 1988; Gregory Stone, 1988; David Touretsky, 1988; computers can process structures. This is not simply assumes some Interlingua Structure of Mental Representations," on this map). Revisionism Andrew Woodfield and Adam Morton, 1988 imagery, and in fact im • age: is just a vague picture-in Symbolic accounts of mind units are primitive Are connectionist There's no tank in that picture. 48 David Braddon-Mitchell and John Fitzpatrick, 1990 Inferential Coherence The ability to make certain inferences is related to the ability to make certain will be exactly correct, after they have been revised on the basis of insights from Cohabitationism Smolensky's treatment of levels is problematic. Smolensky's account of the conceptual, subconceptual, and neural levels (and the relations between them) is problematic. • There are better ways to articulate the levels distinction (Chadrasekaran et. neural level some would deny the claim that images can be represented as filled disputed by the-head metaphor, because operations on these data structures without explaining them. An operational approach is superior Verbal Interlingua Visual - 70 issue areas in the 7 maps Systematicity can be explained by natural selection. Systematicity others. For example, an organism that can infer p from p & q can also infer q is revised by cells in an array. (such as rotation and can be explained by natural selection rather than by a specific architecture such connectionism. Connectionist and symbolic architectures because it produces networks vulnerable from p & q. This is also called systematicity of inference. Proponents: This position al., Quarnton, and Woodfield and Morton). However, the imagery scanning) can be factual information • The goal of cognitive science is not conceptual and neural levels with is as the language of thought. An explanation based on natural selection avoids the cohabitate the mind. Connectionist has been articulated by networks perform low-level perceptual debate in psychology is simulated on a computer. rather than mere formal disputed need for supplementary evidence that must be provided to support a specific subsymbols in between, but rather a "golden age" in which a thoroughly relevant to the is 110 Allan Paivio, 1971 by architectural hypothesis and helps explain how the mind develops over time. Productivity several authors but has not and motor tasks, which interface with the models. The dual code theory. Knowledge can be encoded using a From a finite stock of resources, a person can think an indefinite number of been explicitly endorsed. symbol processor of the mind. understood neuroscience informs a thoroughly understood cognitive computational issue, so disputed Hubert Dreyfus psychology (Lloyd). 95 Zenon Pylyshyn, 1973 dual code of images and associated verbal strings. Whereas - 32 sidebars history and further it is represented here. 92 Stephen Kosslyn and by • The three-level distinction is too simple. There are more levels and Proponents: John Barnden and Walter to the arguments thoughts (or say an indefinite number of things) by combining atomic and neither images nor verbal strings will by themselves provide Rose ity That's a problematic Images are not primitive James Pomerantz, 10 Marvin Minsky and 11 David Rumelhart and atic molecular representations in various ways. This is also known as generativity. Schneider. distinction! objective representation of the world, they can be combined stem Hybridism modeling strategies than just 3 (Quarnton). explanatory concepts. To be 1977 Seymour Papert, 1969, James McClelland, Sy Cognitive researchers should develop hybrid • The treatment of levels is eliminativist (Rey, Schneider, Touretsky). explanatory, images must play a Pylyshyn uses the 93 Zenon Pylyshyn, into coherent representations that do represent the world. as articulated by 1986b John Loves Sally and Sally Loves John models that incorporate aspects of symbolic • The treatment of levels is implementationalist (Hanson). role in causal explanations. But wrong notion of an 1981 David Rumelhart, Multilayer perceptrons • The three-level distinction is incoherent (Antony and Levine). The picture-in-the- against physical and connectionist architectures. This position the mere experience of imagery image. Pylyshyn background • There should be more focus on the neural level (Lloyd, Mortenson, Rueckl). James McClelland, and FARG, can compute all is usually taken as a practical approach to is supported by 94 Allan Paivio, 1971 give us no reason to believe that attacks an extreme head metaphor 1986b relevant functions. Sally Loves John and John Loves Sally modeling, not as a philosophical standpoint. Ecumenicalism • The relationship between the subconceptual and conceptual levels is not Images are primitive is images play such a causal role. "picture-in-the-head" covertly influences 111 Stephen Kosslyn and James Pomerantz, 1977 112 Zenon Pylyshyn, 1973 It is necessary to incorporate is One-layer perceptrons can't compute certain The limitations described by Minsky and Papert do Proponents: Stan Kwasney and Kannaan Faisal, Trent Lange, and Michael Dyer. "everything that works" as we ℘ one of approximation, but of part–whole (Bechtel). • The analogy between Smolensky's levels and Newtonian and quantum functional components of disputed by Images in themselves (prior to interpretation) are epiphenomena theory of images, according to which they is disputed the image theorists. Even if no Quasi-pictorial images can do the same work as symbolic descriptions. An adequate theory of Images cannot encode symbol systems? functions. One-layer perceptrons not apply to multilayer develop theories of mind. thought. Thought by one takes the picture- mental activity can be formed without assuming an knowledge. Knowledge consists of disputed is physics is flawed (Cleland). that ride above a causal substrate are like mental • There should be closer contact between levels than Smolenksy cannot, in principle, compute networks. Minsky and Connectionism, symbolicism, cannot be explained in-the-head metaphor underlying propositional deep structure. information that applies to a range of by certain functions, including: • parity—whether an odd or disputed by Papert ignored such 49 Brian McLaughlin, 1993b neuroscience, and perhaps ℘ suggests (Rueckl). simply in terms of verbal of propositions, like foam rides atop a wave. photographs. But such mental photographs seriously, the metaphor • Knowledge can be gleaned from images in the same can do the is possible situations. Analogue images, Natural selection does not explain • Levels are nothing but pragmatic constructs (Stone). networks in part because other approaches will stimulus-and-response is implicitly drawn on way that it can from be gleaned from sense however, only carry information about Note: Pylyshyn argues elsewhere would lack intrinsic same as disputed even number of units in the there were no useful 50 David Chalmers, 1993 systematicity. Natural selection can't explain contribute to our patterns. The that, in general, conscious when image theorists perceptions. by the situations from which they arose; structure. Images, unlike input layer are active • connectedness—whether all learning procedures for training them. Today, is Connectionist implementations of classical machines possess compositional semantics. Because connectionist implementations of classical systematicity, because natural selection doesn't describe its constitutive bases. Evolutionary understanding of how the mind operates. Ecumenicalism ℘ 68 Paul Smolensky, 1988b phenomena of meaning, memory, learning, and processes lack explanatory power. photographs, are richly appeal to the spatial properties of images. • Representations using images may require less storage and may be more efficient than those that in themselves they lack generality of application. The argumentation maps: 3 John Searle, 1990b; Roger Penrose, 1989 disputed is theory can describe the historical origins of has been called "theoretical structured entities that Note: Also, see the "Can computers active units in the input layer however, there are several by machines have a compositional semantics (see sidebar, "Constituent Structure of The extremist fallacy. A common problem with the various attacks on language involve a still have some spatial use propositions. Rose Connectionist networks are formal systems. Any function that can be computed disputed constitutive bases, but it can only do so after the is represent the analogue properties of on a connectionist network can also be computed on a serial machine. In fact, most current connectionist networks are simulated on serial machines. Conversely, connectionist networks are connected to all other active units (either directly, effective algorithms for training multilayer Mental Representations," on this map), Fodor and Pylyshyn's claim that no connectionist model could have a compositional semantics must be false. This by constitutive bases themselves have been described. So, evolutionary theory by itself isn't pluralism" by William James. Proponents: Eric Dietrich and ? ℘ disputed by the 3-level distinction is that they commit the following "extremist fallacy": There are only 2 positions on the connectionism/symbolism issue: coordination of images with words that is not properties. Supported by "Images Are Quasi-Pictorial Representations," Box 76. images?" arguments on this map. or via other units that are networks. structural flaw in Fodor and Pylyshyn's argument can be traced to their Chris Field, Robert van Gulick, Jay eliminativism and implementationalism. Any view that rejects one fully explained by the can be used to implement classical serial processing. Thus, arguments directed against the underappreciation of the difference between local and distributed representations. enough. Rosenberg, and Gregory Stone. behaviorist program. is formal character of symbol manipulators apply equally well against connectionist networks. active). must embrace the other; if it embraces both, it is incoherent. - arrange debate so that the cur- disputed by Supported by "The Chinese Room Argument," Map 4, Box 3; "Mathematical Insight is Non-Algorithmic," Map 7, Box 23. is disputed 12 Tim van Gelder and Robert Port, 1995 Connectionists fall into a computational mindset. Connectionists take con • sti • tu • tive bas • es: Capacities that work together to constitute a ℘ But the subsymbolic paradigm rejects both eliminativism and implementationalism, forging a "limitivist" middle road. Note: Smolensky separately addresses the various other attacks on his 97 Zenon Pylyshyn, 1981 Images are cognitively cog • ni • tive • ly pen • e • tra • ble: A mental Can computers represent by higher-order capacity. For example, the Limitivism some steps in the right direction, but they fall into a computationalist mindset, Eliminativism treatment of levels. penetrable. Images are phenomenon is cognitively penetrable if it can be for example, when they substitute activation patterns for symbols. Connectionists Representational capacities to have the beliefs that "John Connectionism and Symbolic processes are accessed and altered by other thought processes. Note: The point that connectionist networks and symbol systems can simulate each other is approximations of lower level cognitively penetrable, in that rent stopping point of each is widely accepted. Searle and Penrose use this point to show that their Chinese Room and should focus more on the dynamical systems approach, which views the mind level (compositional Dog Cha Cat ses Sally loves Sally" and "Sally loves John" are neuroscience capture all they can be altered in various For example, a belief that the cat is on the mat is Postulates of the Subsymbolic Paradigm ℘ disputed Loves John both constitutive bases of the higher-order subsymbolic (connectionist) is cognitively penetrable, because it can be altered by Gödel arguments, respectively, apply to connectionist networks as well as to classical AI as a complex system that evolves through time. Connectionists need to "take by semantics) important aspects of ways by what a subject thinks. capacity to have systematically related activity, which is an abstraction is supported by further information that the cat is actually a dummy. the analogue properties systems. the leap out of the computational mindset and into time" (p. 3). mind. High-level, 1. Subsymbols are fine-grained constituents of symbols. disputed For example, in a biological beliefs about John, Sally, and love. symbolic accounts of the from biological processes in the by brain. context an image of a rose is According to Zenon Pylyshyn, those phenomena mind should be ℘ Proponent: Paul Smolensky. ℘ 2. The relation between symbolic models (studied by classical AI) and subsymbolic models seen as a composition of that are cognitively penetrable are generally explained debate thread is easily seen Implementational, eliminated from (studied by connectionism) is similar to the relation between classical Newtonian physics and petals, sepals, leaves, and cognitive science. quantum physics. Symbolic models are rough, macro-level approximations of subsymbolic stems, whereas in an artistic in terms of symbolic rule-governed processes that 4 Jack Copeland, 1993 of images? is supported by 3 Invalid Inferences connectionist level operate on the more basic, cognitively impenetrable is disputed Simulations of connectionist networks are not duplications. It Searle argues But that implies that Thus, Searle's (no semantics) Connectionist Representations Proponents: Stephen Stitch, Patricia ℘ activity. Images may not be primitive, but context the image is seen as a composition of color patches, components. 114 Mark Rollins, 1989 Analogue images have an internal by is invalid to argue from the fact that a that this is invalid the contrapositive connectionism Churchland, and Paul 3. Subsymbolic processes are abstract simplifications of neural processes in the brain (see sidebar, they're still 96 Stephen Kosslyn and shadings, and edges. serial simulation of a connectionist reasoning: form is also invalid: argument is also Like AI researchers, connectionists are concerned with the issue of how Churchland. "Connectionism and the Brain," on this map). At present, it is an open question exactly how James Pomerantz, 1977 Consequently, images cannot syntactic structure. Analogue images are - identify original arguments by 52 Michael Antony, 1991 51 Keith Butler, 1993a mental states represent the world. Several kinds of connectionist important. "structured configurations" whose syntactic network can't think to the conclusion invalid. Neural Eliminativism subsymbolic models relate to neural models. Images are important even if be explanatory primitives; Systematicity is not enough to There are no semantics at the connectionist representations are commonly distinguished. they're not primitive explanatory concepts. is structure can be modeled on a computer. They that an actual connectionist network 1. An actual storm 1. A simulation of a 1. A serial simulation argue for classicism. Even if level. At the relevant level of analysis (the cognitive The only relevant level of they break down into simpler cog • ni • tive • ly im • pen • e • tra • ble: A disputed have this syntactic structure by virtue of their 5 John Searle, 1990b can't think. In fact, the invalidity can description of the mind is 4. Conscious application of rules takes place in a conscious rule-interpreter. The conscious rule It may be true that images are not primitive, low-level parts depending on context. mental phenomenon is cognitively impenetrable if it can make us wet. storm cannot of a connectionist classicism provides the best account and compositional level), connectionist No entities. However, explanation does not always require 113 by relations to prototypes and schemas. The Chinese Gym argument. Suppose the basic Chinese Room were expanded to a large be demonstrated by modifying one of at the neural level. Even interpreter interprets rules sequentially, and is relatively slow. It is well-suited to novel cannot be accessed or altered by other thought processes. gym full of monolingual English-speaking men. The men are spread out like the nodes in a 2. Therefore, a make us wet. network can't think. is of systematicity, it does not follow implementations of classical models are classical, not Local Representations symbols, the connectionist and information, as well as to consciously formulated rules and knowledge. As such, it is best an understanding of what happens at the lowest Computers cannot Searle's own arguments about simulation For example, a person's perception of red is cognitively over 380 protagonists world- network, and they follow English rule books that tell them what symbols to pass to one another. simulation of a 2. Therefore an 2. Therefore, an actual disputed that the mind has a classical connectionist. At the implementational (connectionist) In local representational just brains! subsymbolic accounts analyzed at the conceptual level as a symbolic process. possible level. For example, you wouldn't learn much represent the analogue (see "Simulations Are Not architecture. Systematicity is just impenetrable, because it can't be changed to a perception properties of images. Through this procedure they carry out the same computations as a connectionist network would Duplications," Map 2, Box 23). storm can make actual storm connectionist network by level, they have no compositional semantics. So, schemes, each node in a should be eliminated. about architectural design just by studying bricks, 115 Zenon Pylyshyn, 1973 of green. to produce Chinese speech, but none of the men understand Chinese. This example shows that instantiating a connectionist network is not enough to produce an understanding of Chinese. us wet. cannot make us wet. can't think. one of many phenomena that a theory of mind must explain (along Chalmers is wrong to claim that connectionist implementations of classical models possess network represents some concept. Dog node Cat node Proponent: Walter Freeman (to some extent). ℘ 5. The conscious rule-interpreter is a virtual machine that is run on an intuitive processor. mortar, steel, and so on. Likewise, the proper level of analysis of perception requires the inclusion of According to Zenon Pylyshyn, cognitively impenetrable Imagery cannot be reduced to discrete computational is disputed Analogue images can't encode knowledge. The analogue interpretation of with other phenomena like perception, compositional semantics, because, to the extent that 6. The intuitive processor is responsible for most behavior, including linguistic behavior, problem images. form without by images makes them too specific to encode Note: Also, see Map 4. In particular, "The Chinese Water Pipe Brain Simulator" (Map 4, Box imagination, emotion, etc.). To show phenomena are primitive explanatory concepts, because is they are connectionist implementations, they have no misrepresenting its wide over 40 years 5) parallels this argument, with the difference that the Chinese water-pipe argument uses one is solving, and all skilled performance. The intuitive processor handles unconscious, learned they remain basic throughout changes of context. disputed knowledge. Knowledge has a generality that disputed that the mind has a classical semantics at all. Distributed Representations activities, operates in parallel, and is relatively fast. The intuitive processor is best analyzed continuous analogue can only be captured by propositions. man instead of a gym full of men. 8 Paul and Patricia Churchland, 1990 by by architecture, then, the classicist must In a distributed representation, Dog pattern at the subconceptual level as a subsymbolic process. Experiment properties. No individual neuron understands Chinese. It is demonstrate that classicism does better a pattern of activity over the of activity Eliminativism Note: Also, see "The Brain than connectionism (or any other whole set of nodes represents 100 Stephen Kosslyn, 1973 is An Analogue Device," irrelevant that no one in Searle's Notes: • Smolensky also calls this argument the "proper treatment of connectionism" or "PTC." competing theory) at accounting for a concept. 2 1 42 Notes: Scanning visual images. Map 3, Box 9. 116 Ned Block, 1983 Chinese Gym understands Chinese. • These positions are discussed by Dinsmore (1992), Smolensky (1988b, pp. 59–62), and Pinker and Prince - make the current frontier of 53 Brian McLaughlin, 1993a • Smolensky distinguishes levels of analysis (conceptual, subconceptual, and neural) from all of the relevant phenomena, not is Subjects were requested to memorize is Digital computers can't process No neuron in an English-speaker's 1 4 is just systematicity. disputed Antony misrepresents the 0 0 00 (1988, pp. 75–78). is supported by a series of pictures and later to disputed disputed analogue images. Digital computers will brain understands English, even cognitive systems (symbolic, subsymbolic, and neural). It may be important to remember that though the brain as a whole does. by argument. Fodor and Pylyshyn do Microfeatural • Few theorists or researchers explicitly position themselves along this spectrum. Many fall into more than one his distinction is often ignored in other authors' discussion of levels. imagine them one at a time. They by by probably never be able to process analogue 6 Jack Copeland, 1993 7 Paul and Patricia Churchland, not argue for a classical architecture Representations of these categories or lie somewhere on the borders between them. were asked to focus on one end of is supported by information such as images. The brain The systems reply to the Chinese Gym. 1990 solely on the basis of systematicity. Another form of the imagined picture and to identify processes imagery by using analogue Searle's Chinese Gym argument commits the The Chinese Gym requires a I don't They acknowledge that a theory of Conceptual Level Subconceptual Level Neural Level a number of its features. It was found debate easily identifiable is is 54 Andy Clark, 1991 representational scheme is physiological mechanisms. In order to do the same fallacy as the Chinese Room argument preposterous number of understand Systematicity is a conceptual cognition must account for other microfeatural. Each node in a Postulates of the Dynamical Approach to Cognition Preferred level of description Preferred level of description Preferred level of description that the amount of time needed for same thing, digital computers will probably disputed disputed commits. It is invalid to infer that the gym as people. Searle's thought English or rather than an empirical law. phenomena as well, and in fact they network represents some low- of symbolic processes. At of subsymbolic processes in the of processing in the brain. The the subjects to correctly respond have to be supplemented by analogue by a whole doesn't understand Chinese from the by Chinese! discuss productivity and inferential Tail- Dog neural level describes many 98 Mark Rollins, 1989 experiment could not reasonably Systematicity fails to argue against level feature of a higher-level ear 1. Natural cognitive systems are dynamical systems. this level, concepts and rules connectionist dynamical system, is supported by corresponded to the time it would 99 Stephen Kosslyn, Steven Pinker, George mechanisms. fact that the individuals in the gym don't be implemented. His gym would coherence as additional phenomena wagging are consciously formulated and the fundamental level for biological details that are not have taken them to scan an actual The cognitive impenetrability condition E. Smith, and Steven P. Schwartz, 1979 the empirical hypotheses of the concept (see "The Coffee node doesn't isolate basic elements. The understand Chinese. have to hold the entire populations that a theory of cognition must Story," Box 35). node 2. The mathematics of dynamical systems provide a general framework for constructing and testing theories and understood. This level is the study of mind. Fine-grained relevant to the subconceptual picture. Conclusion: internal images Cognitive impenetrability does not connectionist architecture, because - provide summaries of eleven Note: Also, see the "Can the Chinese Room, of more than 10,000 Earths. explain. similar to Newtonian physics features of symbols and symbol level, which makes appropriate have spatial properties analogous to requirement that basic elements of functional argue against image theory. The systematicity is a conceptual law of cognition. architecture must be cognitively impenetrable is considered as a total system, think?" arguments is in that it offers an processing are studied at this computational abstractions. those of external images. cognitive impenetrability condition does disputed that deals with holistic thought flawed. on Map 4. approximation of the lower- level. not refute the image theory because the by ascriptions rather than with is 55 Brian McLaughlin, 1993b Patterns of Activity 3. A dynamical system is a set of changing aspects of the world, represented by variables. "It is, in short, of level microtheory. • Many functions of the body, such as digestion image theory does not assume that all image in-the-head thought processes. disputed It is usually assumed that connectionist representations correspond to the essence of dynamical models of this kind to describe how processes unfold, moment by moment in real are cognitively penetrable but still basic to Other imagery arguments That is, systematicity tells us about by The challenge stands. If patterns of activity (individual activity values in local representations, time" (van Gelder and Port, 1995, p. 19). is processes are cognitively impenetrable. The many explanations. major philosophical camps of systematicity is a conceptual truth image theory acknowledges that some how we judge whether other about cognition, as Clark says, then it distributed activity values in distributed representations). As a result, John Loves Sally Experiment disputed • The cognitive impenetrability condition image processes are cognitively penetrable people are thinking; it does not tell connectionist representations are constantly changing while a network 4. A state of the system is the way the variables happen to be at a given point in time. by Are connectionist networks is a necessary condition for the 103 Stephen Kosslyn, 1994 assumes a fixed level for elements in the whereas others are not. The theory also 15 Ronald Rosenfield, David us about what is going on inside is runs, and are thus highly context sensitive. These are sometimes called functional architecture. But functional is supported by Touretsky, and the their heads. adequacy of a cognitive theory that it "active" representations. 5. The state space of a system is the set of all states the system might be in. Mental rotation. Subjects were shown an object and were recognizes that determining which are disputed explain systematicity. Therefore, later presented with a rotated version of the same object. The elements sometimes change levels in the which is an important research issue. Boltzmann Research by systematicity still poses a challenge Connectionism and the Brain further the object was rotated from its initial position, the hierarchy. is supported by 117 Hubert Dreyfus and Stuart Dreyfus, 1986 the protagonists (or schools of Group, 1988 Weight Representations 6. The behavior of a dynamical system is governed by differential equations. The differential equations to connectionism. longer it took the subjects to decide if the figure was the same Computers can't use images unless they transform them into descriptions. If a computer is like human neural networks? Simplicity has practical Sometimes the weights in a connectionist network are taken to represent describe how the state of the system changes continuously through time. Connectionist networks are usually described as being "neurally plausible" or "neurally inspired." the world. Weight representations come to reflect aspects of the world as They do not exactly simulate the operation of the brain, but they do capture some of its basic as the original one. The subjects rotated images in their heads 101 Charles L. Richman, David B. Mitchell, and J. Steven Reznick, 1979, as articulated by Zenon Pylyshyn, 1981 going to make inferences from an image, then the image must first be decomposed into a list of facts. Humans advantages over are not constrained in this way; they can work directly with images. biological accuracy. the network learns. Weight representations change much more slowly 7. Dynamical systems exhibit such features as attractors, limit cycles, complexity, bifurcation, and chaos. computational principles. For example: in much the same way that they would rotate physical objects Evidence from scanning and rotation experiments is problematic. Kosslyn's experimental results may Skarda and Freeman fail to than patterns of activity do, and are sometimes called "passive" Many of these features can be visualized—to a point—in graphic presentations. • Both brains and connectionist networks process information by using numerous interconnected in space. Conclusion: the mental image being rotated has be invalid for the following reasons. Note: Also, see sidebar, "Postulates of Dreideggereanism," on Map 3. • Subjects' tacit knowledge of real objects makes them think that they are supposed to work with images in the same is show that their dynamic representations (see "Weight Representations Avoid the Regress," Box processing units (neurons in the brain; nodes in a connectionist network). spatial properties analogous to those of a object. • Both brains and connectionist networks learn by modifying connections between processing units disputed thought). 56 Keith Butler, 1993b Note: Kosslyn credits R. N. Shepard, J. Metzler, and L. A. way that they work with real objects. As a result, subjects are not literally scanning or rotating images, but are 14 Christine Skarda and models have any real The claim that systematicity is a conceptual fact is unsupported. The reasons Clark 37). Note: The dynamical system approach has been applied to many aspects of mind, including development, by Walter Freeman, 1987 is advantages. Unmapped Territory language, perception, action, and the brain. Proponents include Walter Freeman, Timothy van Gelder, Christine (synapses in the brain; weights in a connectionist network). Cooper with earlier versions of this experiment. mentally simulating a real scanning or rotation of a real object. That is, the task demands of the experiment confound 118 Hubert Dreyfus and Stuart Dreyfus, 1986 • They leave the role of chaos gives to support the claim that systematicity is a conceptual fact are unconvincing. • Both brains and connectionist networks exhibit distributed, "self-organizing" behavior. How do I get a disputed • Thought ascriptions are not necessarily holistic, because "organisms can have individual thoughts 13 is Connectionist network are too simple. Real neural networks, Skarda, and Robert Port. Other notable proponents include Christopher Zeeman, Jean Petitot, and Rene Thom. its results. Computers can't recognize similarities between whole images. The connectionist biological assumption. disputed observed in living organisms such as rabbits, exhibit complex by vague. Additional In some sense all cognitive scientists are dynamicists, to the extent that they accept contemporary mathematics • The experimenter is able to affect the subject with "nonverbal cues, tacit messages, ... loaded answers to questions, is supported by Humans directly recognize similarities between images. For example, involving a host of concepts in the absence of evidence for the versatile deployment of those Connectionist networks are similar to real neural by dynamics and chaotic activity that connectionist networks lack. • They argue ineffectively concepts" (p. 39). connectionist with its dynamical formalisms. Despite such similarities, connectionist networks make various abstractions (or "simplifying and so on" (p. 544). This gives rise to a variety of undesirable experimenter effects. a human can directly perceive 2 different faces as being gentle, mocking, representations • Thought ascriptions do seem to have something to do with in-the-head mental processing. networks. These complex dynamics seem too messy from an engineering that pattern completion assumptions") from real neural processes in the brain. For example: or puzzled. Computers, by contrast, must compare images by assigning Note: Also, see the "Is the brain a computer?" standpoint but are critical to an understanding of neural dynamics. differs from the activity of arguments These postulates are adapted from Timothy van Gelder and Robert Port (1995). For a visual discussion of • Connectionist networks use fewer units and connections than are found in real neural networks. 102 Stephen Kosslyn, Steven Pinker, George E. Smith, and Steven P. Schwartz, 1979 them features and then comparing those features using some objective arguments on Map 1 and the "Is the relation between So, it appears that systematicity is an empirical (and not a conceptual) issue, and whether connectionists It is usually assumed, for example, that a given node corresponds to a large collection of real real networks. can account for it remains an open question. dynamics, see Ralph Abraham and Christopher Shaw (1982). Alleged problems with the imagery experiments have been disconfirmed. criterion. But it is not clear how the perception of 2 faces as gentle or hardware and software similar to that between human • Although it is true that neurons. mocking involves the recognition of shared objective features. • Connectionist networks use weights that can switch between inhibitory and excitatory. Synapses is Task demands and experimenter effects do not refute scanning and rotation experiments. brains and minds?" arguments on Map 3. One should be feedback is important, there disputed • The claim that task demands caused subjects to simulate how they would behave with 16 Jerry Fodor and Zenon Pylyshyn, 1988 deeply is no point in "blindly in the brain are either inhibitory or excitatory, but never both. • Many connectionist learning schemes, especially backpropogation, are clearly nonbiological. by real objects is disconfirmed by experiments that make no mention of physical motion. "Brain-style" modeling can be suspicious of the simulating neural circuitry." References to "physical motion" and "scanning" are replaced by such phrases as "glance misleading. Basing psychological theories on heroic sort of Connectionists accept Nothing in the brain resembles the process of backwards propagation of error that is used in most up" and "shift attention." is similar to facts about the brain can be misleading. Although brain modeling simplifications in order to get modern connectionist networks. • The claim that subjects give answers they think the experimenters want to hear is 24 Steven Pinker and Alan Prince, 1988 printed copy? neural inspiration seems useful, it has led to the that purports to useful scientific results. 25 Mark Seidenberg, 1992 26 Mark Seidenberg, 1992 29 Terence Horgan and John Tienson, 1991 is disconfirmed by the fact that subjects often give responses that the experimenters didn't is revival of such weak psychological theories as: address the The past-tense model does not argue against rule- Any system can be described by The multiple realizability defense. It To what extent connectionist networks will ultimately come to resemble the brain is an open question disputed anticipate. These responses could not have been suggested by the experimenters. Connectionist models of disputed • associationism problems of based explanation. The past-tense model is problematic a dual-route model. Pinker and language do not is not possible to formulate rules for and a topic of ongoing research. The table that follows summarizes some of the more detailed by by • microfeature analysis cognition (p. 64). 18 Nick Chater and Mike Oaksford, 1990 in numerous ways. For example: Prince postulate a dual-route model of implement classical connectionist representations (or comparisons between connectionist networks and the brain. • statistically based learning Fodor and Pylyshyn Biology is relevant to a theory of cognition. Low-level • It cannot represent certain words. linguistic knowledge, according to models. Connectionist models representation instantiation rules), because They're all properties can constrain high-level properties, even if the 2 levels • It cannot learn certain rules. which the ability to transform verbs into of language, such as Seidenberg such representations can be realized by 0 0 0 dog to me. Christine Skarda are structurally dissimilar. For example, physics constrains chemistry, • It learns rules found in no human language. the past tense consists in either Paul Smolensky, 1988b 17 Jerry Fodor and Zenon Pylyshyn, 1988 Facts about the brain may be irrelevant to facts about thinking. is even though processes at the 2 levels are structurally different. Similarly, biology constrains a theory of cognition. • It fails at its assigned task ofmastering the past tense of English. is disputed • following a rule (add "-ed" to the verb; e.g., "guide" becomes "guided"), or is and McClelland (1989), do not implement a dual-route model, because the models cannot be multiple node-level descriptions. A representation of "dog," for example, could be realized in a variety of different (though 0 .1 0 .2 0 .1 and Walter Freeman, 1987 104 John Anderson, Postulates of Image Psychology Legend Structures at different levels of organization are often dissimilar. For disputed To overcome such difficulties, connectionism will have to by • consulting a list of exceptions (e.g., disputed decomposed into separate similar) node-level activation patterns. Thus, .2 .2 .2 1978 1. Verbal processing cannot explain all cognitive processing. is implement certain features of rule-based, symbolic theories. "run" becomes "ran," "weep" becomes by example, rocks and rivers have little in common with the atoms they are by systems for applying rules and no single node-level description can fully 0 0 0 The empirical Imagery is necessary as well. Focus Box: The lowest-numbered box in each issue area is an introductory focus box. disputed constructed of. Similarly, thinking may have little in common with the So, connectionist models are either inadequate as models of "wept," etc.). handling exceptions. Moreover, describe the "dog" representation. Moreover, evidence is The arguments on these maps are organized by links that carry a range of meanings: The focus box introduces and summarizes the core dispute of each issue area, sometimes by Physics language, or at best offer an implementation of classical rule- But any system can be described in this .5 .5 .6 inconclusive. It Rose 2. Imagery and verbal processing are alternative coding systems, neural structures it is implemented in. So, we should be careful about Chemistry they behave in behaviorally node-level discrepancies make it possible for Brains Connectionist Brains (observed Connectionist basing a theory of cognitive architecture on a theory about the brain. 19 Keith Butler, 1993a based accounts. way, because a set of rules only has to plausible ways not predicted by the "dog" representation to produce a variety cannot be decided on or "modes of symbolic representation." as an assumption and sometimes as a general claim with no particular author. Functional aspects of implementation are Note: Compare the structure of this argument to that of "The fit some cases, with the rest being treated the dual-route model. of different behaviors. have these networks have in rabbits) have networks have the basis of behavioral Arguments that uphold or defend another claim. Examples include: relevant. It is true that material properties of the brain Connectionist Dilemma," Box 31. as exceptions. Such an approach is "like properties: these properties: these properties: these properties: evidence whether 3. Behaviorism puts emphasis on linguistic phenomena because is are irrelevant to psychological theory construction. constrains saying that all of the observations in my ? is supported by supporting evidence, further argumentation, thought experiments, Arguments with No Authors: Arguments that are not attributable to a particular source disputed pictorial or it claims that only verbal reports are empirically accessible. is not by However, connectionism is not concerned with material experiment fit a particular hypothesis Locally dense propositional However, experiments can be designed that make mental extensions or qualifications, and implemented models. (e.g., general philosophical positions, broad concepts, common tests in artificial properties. Connectionism is concerned with functional except for the ones that I have decided Neurons are located Nodes have no Typically no similar to is feedback, which feedback representations offer an imagery empirically accessible as well. intelligence) are listed with no accompanying author. properties of the brain, like graceful degradation and to exclude" (p. 94). disputed in two- and one- spatial location. provides for a adequate explanation of parallel processing, which are relevant to cognition. is dimensional space. "continuum of local mechanisms by mental imagery. What 4. The more concrete or "thing-like" a stimulus is, the more disputed interactions" the experimental likely it will be associated with an image, rather than with A charge made against another claim. Examples include: Citations: Complete bibliographic citations can be found in the booklet that accompanies Follow a rule by Synapses are located Connections between is evidence does support a verbal, process. disputed logical negations, counterexamples, attacks on an argument's this map. Input in three-dimensional nodes have no spatial Complex mixtures of is a theory about the Consult a list of space. location. excitatory and At best, layer-to- by emphasis, potential dangers an argument might raise, thought 28 Anticipated by Terence Horgan and John Tienson, 1991 You can order artist/researcher process through which 5. Imagery is a parallel processing system that stores and exceptions The syntactic argument. The representation-without- inhibitory feedback layer feedback representations are manipulates spatial information. Verbal processing, by experiments, and implemented models. Anticipated by Where this phrase appears in a box, it identifies a potential Do connectionist rules conception of connectionism cannot succeed, because Location of Connections between employed. contrast, specializes in serial tasks. 21 Andy Clark, 1992 22 David Rumelhart and 30 Kenneth Aizawa, 1994 nodes have no spatial Chaotic and attack on a previous argument that is raised by the author Explicit rules are necessary in plastic domains. In order to perform well in James McClelland, 1986 rules can always be formulated to describe a network's Variable outputs can be described by rules. synapses strongly High-level chaotic Note: Anderson Steven Pinker Alan Prince affects signal location. activity that puts the oscillatory so that it can be disputed. changing circumstances, or plastic domains, a cognitive system must have access to Regularity without rules. representation-level processing. Such rules can Even granting that a given state of a connectionist supports this argument 6. Chains of symbolic transformation can involve images, • specify how the parts of a representation are instantiated activity are not is interpreted as signed copies of all seven maps Connectionist networks exhibit lawful network could lead to a variety of further states, interactions. network into an "I with a formal proof. words, or both. These chains mediate perception, learning, explicit rules that it can manipulate and redescribe. For example, in certain circumstances modeled because A distinctive reconfiguration of an earlier claim. 27 Terence Horgan in nodes and connections (representation instantiation don't know state," As articulated by Where this phrase appears in a box, it identifies a reform- networks follow a scientist might want to redescribe Ohm's law (V = C x R), inverting the relationship behavior without following explicit a set of rules could still be formulated to describe Neurons have dense Nodes have uniformly they are undesirable memory, and language. between voltage and resistance. But a connectionist network can't perform such a rules. Regularities emerge from the and John Tienson, 1991 rules), and can its behavior. This is because rule forms have connectivity to dense connections. which allows it to is Representations without rules. • characterize the operation of individual nodes in the network avoid old patterns and from an engineering is 7. Basic concepts like "image," "mediation," "word," ulation of another author's argument. The reformulation redescription, because it lacks an explicit representation of the law. At best, the interactions of low-level processing "implicit conventions for simplification when a nearby neurons. standpoint disputed disputed network can be subjected to extensive retraining. units, rather than from the application is To adequately model cognition, (node-level rules). single input representation might lead to distinct acquire new ones "processing," and so forth should be defined operationally is different enough from the original author's wording to from www.macrovu.com for by Projections between by Note: In making this claim Clark draws on the developmental of high-level rules. Although it may disputed connectionist networks must exhibit is output representations" (p. 484). So, the syntactic Projections between in order to give them precise experimental significance. warrant the use of the tag. This phrase is also used when rules? by representational structure without Node-level rule areas have an node pools have a Destabilizations, Pattern completion psychology of Annette Karmiloff-Smith. be possible to characterize a network's disputed argument holds. devices, which take Unmapped Territory This icon indicates areas of argument that lie on or near the the original argument is impossible to locate other than in following "obligatory, hard, intricate topology. simple topology. which lead an animal These postulates are adapted from Allan Paivio (1971). Other behavior according to explicit rules, by Note: The debate between Aizawa and Horgan a partial input and boundaries of the central issue areas mapped on these maps. none are involved in its underlying Implemented Model representation-level" rules. This new and Tienson about rules is carried out with much through a trajectory proponents (whose opinions may differ on specific points) Additional its articulation by a later author (e.g., word of mouth), or paradigm for cognitive science must There is intricate There is simple, linear of actions then recast the It marks regions of potential interest for future mapmakers to denote a general philosophical position that is given a $500.00 plus shipping and han- mechanisms. more precision than could be captured here. In pattern as a whole include Rudolph Arnheim, Bergen Bugelski, Gordon Bower, arguments 23 David Rumelhart and James McClelland, 1986 include "representations with signal integration in signal integration Roger Shepard, Arthur Staats, and others. Stephen Kosslyn's 20 is supported by Note: This claim is nearly identical David Rumelhart James McClelland The past-tense acquisition model. This network was trained to convert English the technical discussion, issues of quasi- a single neuron. between nodes. 105 Stephen Kosslyn, Steven Pinker, George E. Smith, and explorers. special articulation by a particular author. to "Explicit Rules Are Unnecessary," complex internal structure that can If weighted input is exceptionless rules, probabilistic laws, and ceteris and Steven P. Schwartz, 1979 work grew out of image psychology. Connectionist networks can think without following rules. phrases into the past tense. For example, the network converts "sip" to "sipped," be related in various ways that go Like humans, connectionist networks exhibit fluid, intelligent behavior without following rigid, explicit rules. In general, they are trained Map 3, Box 52. greater than .5, output 1. paribus rules are raised. There is a single signal Note: Skarda and Freeman offer these The empirical evidence favors image "hug" to "hugged," "run" to "ran," and so forth. Although the network's performance beyond mere association; and There are numerous observations as criteria that could lead to more psychology. Image psychology predicts image to exhibit intelligent behavior rather than being programmed with rules. op • er • a • tion • al def • i • ni • tion: A definition of a dling. can be described by rules, no actual rules are utilized in its processing. processing that depends on this signal types. type. Notes: flexible and realistic connectionist models, not rotation and scanning in addition to explaining them. • Also, see the "Do humans use rules as physical symbol systems do?" arguments on Map 3. Note: More recent work along these lines includes that of Mark Seidenberg and structure and those relations. But, as essential limitations on connectionism. Propositional theory, by contrast, does not predict concept in terms of a repeatable operation. For example, anger One of 7 in this Issue Mapping™ series—Get the rest! can be operationally defined in terms of the number of times • Terence Horgan and John Tienson (1991) provide a full characterization of classical and connectionist notions of rule following Fred is supported by James McClelland (1989). for large and significant areas of image rotation and scanning. At best, a propositional cognition there must be no rules Note: Connectionist researchers have a spectrum of concerns. Whereas computational neuroscientists theory can be patched together to explain those a subject hits a dummy in a controlled environment. Also, see is supported by Adams, Kenneth Aizawa, and Gary Fuller (1992) provide a detailed treatment of the relation between different kinds of rules, arguing The remaining 6 maps in this Issue Mapping™ series can be ordered with MasterCard, the "Is the test, behaviorally or operationally construed, a adverting to representational are directly concerned with modeling the brain, neural engineers apply connectionist principles phenomena, but only in an ad hoc way that could just A further discussion of argumentation analysis methodology can be found in the that the differences between them should not be overstated. is supported by structure or content" (p. 248–49). VISA, check, or money order. Order by phone (206–780–9612), by fax (206–842–0296), without concern for neural realism. A similar distinction is made in the field of artificial intelligence, as well explain opposite results. legitimate intelligence test?" arguments on Map 2. © 1998 R. E. Horn. All rights reserved. or through the mail (Box 366, 321 High School Rd. NE, Bainbridge Island, WA 98110). booklet that accompanies this map. Version 1.0 between simulated thinking and "whatever works."