{"title": "On the Computational Utility of Consciousness", "book": "Advances in Neural Information Processing Systems", "page_first": 11, "page_last": 18, "abstract": null, "full_text": "On the Computational Utility of \n\nConsciousness \n\nDonald W. Mathis and Michael C. Mozer \nmathis@cs.colorado.edu, mozer@cs.colorado.edu \n\nDepartment of Computer Science and Institute of Cognitive Science \n\nUniversity of Colorado, Boulder \n\nBoulder, CO 80309-0430 \n\nAbstract \n\nWe propose a computational framework for understanding and \nmodeling human consciousness. This framework integrates many \nexisting theoretical perspectives, yet is sufficiently concrete to allow \nsimulation experiments. We do not attempt to explain qualia (sub(cid:173)\njective experience), but instead ask what differences exist within \nthe cognitive information processing system when a person is con(cid:173)\nscious of mentally-represented information versus when that infor(cid:173)\nmation is unconscious. The central idea we explore is that the con(cid:173)\ntents of consciousness correspond to temporally persistent states \nin a network of computational modules. Three simulations are de(cid:173)\nscribed illustrating that the behavior of persistent states in the \nmodels corresponds roughly to the behavior of conscious states \npeople experience when performing similar tasks. Our simulations \nshow that periodic settling to persistent (i.e., conscious) states im(cid:173)\nproves performance by cleaning up inaccuracies and noise, forcing \ndecisions, and helping keep the system on track toward a solution. \n\n1 \n\nINTRODUCTION \n\nWe propose a computational framework for understanding and modeling conscious(cid:173)\nness. Though our ultimate goal is to explain psychological and brain imaging data \nwith our theory, and to make testable predictions, here we simply present the frame(cid:173)\nwork in the context of previous experimental and theoretical work, and argue that \n\n\f12 \n\nDonald Mathis, Michael C. Mozer \n\nit is sensible from a computational perspective. We do not attempt to explain \nqualia-subjective experience and feelings. It is not clear that qualia are amenable \nto scientific investigation. Rather, our aim is to understand the mechanisms under(cid:173)\nlying awareness, and their role in cognition. We address three key questions: \n\n\u2022 What are the preconditions for a mental representation to reach consciousness? \n\n\u2022 What are the computational consequences of a representation reaching conscious(cid:173)\nness? Does a conscious state affect processing differently than an unconscious state? \n\n\u2022 What is the computational utility of consciousness? That is, what is the compu(cid:173)\ntational role of the mechanism(s) underlying consciousness? \n\n2 THEORETICAL FRAMEWORK \n\nModular Cognitive Architecture. We propose that the human cognitive ar(cid:173)\nchitecture consists of a set of functionally specialized computational modules (e.g., \nFodor, 1983). We imagine the modules to be organized at a somewhat coarse level \nand to implement processes such as visual object recognition, visual word-form \nrecognition, auditory word and sound recognition, computation of spatial relation(cid:173)\nships, activation of semantic representations of words, sentences, and visual objects, \nconstruction of motor plans, etc. Cognitive behaviors require the coordination of \nmany modules. For example, functional brain imaging studies indicate that there \nare several brain areas used for different subtasks during cognitive tasks such as \nword recognition (Posner & Carr, 1992). \n\nModules Have Mapping And Cleanup Processes. We propose that mod(cid:173)\nules perform an associative memory function in their domain, and operate via a \ntwo-stage process: a fast, essentially feedforward input-output mapping1 followed \nby a slower relaxation search (Figure 1). The computational justification for this \ntwo stage process is as follows. We assume that, in general, the output space of a \nmodule can represent a large number of states relative to the number of states that \nare meaningful or well formed-i.e., states that are interpretable by other modules \nor (for output modules) that correspond to sensible motor primitives. If we know \nwhich representations are well-formed, we can tolerate an inaccurate feedforward \nmapping, and \"clean up\" noise in the output by constraining it to be one of the \nwell-formed states. This is the purpose of the relaxation step: to clean up the \noutput of the feedforward step, resulting in a well-formed state. The cleanup pro(cid:173)\ncess knows nothing about which output state is the best response to the input; \nit acts solely to enforce well-formedness. Similar architectures have been used re(cid:173)\ncently to model various neuropsychological data (Hinton & Shallice, 1991; Mozer & \nBehrmann, 1990; Plaut & Shallice, 1993). The empirical motivation for identifying \nconsciousness with the results of relaxation search comes from studies indicating \nthat the contents of consciousness tend to be coherent, or well-formed (e.g., Baars, \n1988; Crick, 1994; Damasio, 1989). \n\nPersistent States Enter Consciousness. In our model, module outputs enter \nconsciousness if they persist for a sufficiently long time. What counts as long enough \n\nlWe do not propose that this process is feedforward at the neural level. Rather, we \nmean that any iterative refinement of the output over time is unimportant and irrelevant. \n\n\fOn the Computational Utility of Consciousness \n\n13 \n\nrelaxation search \n\n.00 \u2022 \u2022 \n\nfeedforward mapping \n\nFigure 1: Modules consist of two components. \n\nis not yet determined, but in order to model specific psychological data, we will be \nrequired to make this issue precise. At that time a specific commitment will need \nto be made, and this commitment must be maintained when modeling further data. \n\nAn important property of our model is that there is no hierarchy of modules with \nrespect to awareness, in contrast to several existing theories that propose that access \nto some particular module (or neural processing area) is required for consciousness \n(e.g., Baars, 1988). Rather, information in any module reaches awareness simply by \npersisting long enough. The persistence hypothesis is consistent with the theoretical \nperspectives of Smolensky (1988), Rumelhart et al (1986), Damasio (1989), Crick \nand Koch (1990), and others. \n\n2.1 WHEN ARE MENTAL STATES CONSCIOUS? \n\nIn our framework, the output of any module will enter consciousness if it persists in \ntime. The persistence of an output state of a module is assured if: (1) it is a point \nattractor of the relaxation search (i.e., a well-formed state), and (2) the inputs to \nthe module are relatively constant, i.e., they continue to be mapped into the same \nattractor basin. \n\nWhile our framework appears to make strong claims about the necessary and suf(cid:173)\nficient conditions for consciousness, without an exact specification of the modules \nforming the cognitive architecture, it is lacking as a rigorous, testable theory. A \ncomplete theory will require not only a specification of the modules, but will also \nhave to avoid arbitrariness in claiming that certain cognitive operations or brain \nregions are modules while other are not. Ultimately, one must identify the neu(cid:173)\nrophysiological and neuroanatomical properties of the brain that determine the \nmodule boundaries (see Crick, 1994, for a promising step in this regard). \n\n3 COMPUTATIONAL UTILITY OF CONSCIOUSNESS \n\nFor the moment, suppose that our framework provides a sensible account of aware(cid:173)\nness phenomena (demonstrating this is the goal of ongoing work.) If one accepts \nthis, and hence the notion that a cleanup process and the resulting persistent states \nare required for awareness, questions about the role of cleanup in the model be(cid:173)\ncome quite interesting because they are equivalent to questions about the role of \nthe mechanism underlying awareness in cognition. One question one might ask is \nwhether there is computational utility to achieving conscious states. That is, does a \nsystem that achieves persistent states perform better than a system that does not? \n\n\f14 \n\nDonald Mathis, Michael C. Mozer \n\nDoes a system that encourages settling to well-formed states perform better than a \nsystem that does not? We now show that the answer to this question is yes. \n\n3.1 ADDITION SIMULATION \n\nTo examine the utility of cleanup, we trained a module to perform a simple multistep \ncognitive task: adding a pair of two-digit numbers in three steps. 2 We tested the \nsystem with and without cleanup and compared the generalization performance. \n\nThe network architecture (Figure 2) consists of a single module. The inputs consist \nof the problem statement and the current partial solution-state. The output is \nan updated solution-state. The module's output feeds back into its input. The \nproblem statement is represented by four pools of units, one for each digit of each \noperand, where each pool uses a local encoding of digits. Partial solution states are \nrepresented by five pools, one for each of the three result digits and one for each of \nthe two carry digits. \n\nProjection \n\nInput-ot,rtpu \nt \nmapping \n\n.......... \n\n( \n\n(copy) ~ if) \n~ ~ jU \n+ \n\nleany tI \n\nI result 31 \n\n1 result 11 \n\nIresult21 \n\nleany21 \n\nI \n\nhidden units \n\nI \n\n(copy) \n\n.,/ \n\n~[~E::~:::][:iIl[iI] Iresult111resutt 211result311eany 111eany 21 \n\n\"-\n\n\" \n\nFigure 2: Network architecture for the addition task \n\nEach addition problem was decomposed into three steps (Figure 3), each describing \na transformation from one partial solution state to the next, and the mapping net \nwas trained perform each transformation individually. \n\n? ? \n\n48 \n+ 62 \n??? \n\nstep 1 \n\n--.. \n\n? 1 \n\n48 \n+ 62 \n??O \n\nstep 2 \n\n--.. \n\n1 1 \n\n48 \n+ 62 \n?10 \n\nstep 3 \n\n--.. \n\n1 1 \n\n48 \n+ 62 \n11 0 \n\nFigure 3: The sequence of steps in an example addition problem \n\nStep 1 Given the problem statement, activate the rightmost result digit and right(cid:173)\n\nmost carry digit (comprising the first partial solution). \n\n2 Of course, we don't believe that there is a brain module dedicated to addition problems. \n\nThis choice was made because addition is an intuitive example of a multistep task. \n\n\fOn the Computational Utility of Consciousness \n\n15 \n\nStep 2 Given the first partial solution, activate the next result and carry digits \n\n(second partial solution). \n\nStep 3 Given the second partial solution, activate the leftmost result digit (final \n\nsolution). \n\nThe set of well-formed states in this domain consists of all possible combinations of \ndigits and \"don't knows\" among the pools (\"don't knows\" are denoted by question \nmarks in Figure 3). Local representations of digits are used within each pool, and \n\"don't knows\" are represented by the state in which no unit is active. Thus, the \nset of well-formed states are those in which either one or no units are active in \neach pool. To make these states attractors of the cleanup net, the connections \nwere hand-wired such that each pool was a winner-take-all pool with an additional \nattractor at the zero state. \n\nTo run the net, a problem statement pattern is clamped on the input units, and \nthe net is allowed to update for 200 iterations. Unit activities were updated using \nan incremental rule approximating continuous dynamics: \n\nai(t) = TI(L: Wijaj(t - 1)) + (1 - T)ai(t - 1) \n\nj \n\nwhere ai(t) is the activity of unit i at time t, T is a time constant in the interval \n[0,1]' and 10 is the usual sigmoid squashing function. \nFigure 4 shows the average generalization performance of networks run with and \nwithout cleanup, as a function of training set size. Note that, in principle, it is \nnot necessary for the system to have a cleanup process to learn the training set \nperfectly, or to generalize perfectly. Thus, it is not simply the case that no solutions \nexist without cleanup. The generalization results were that for any size training set, \npercent correct on the generalization set is always better with cleanup than without. \nThis indicates that although the mapping network often generalizes incorrectly, the \noutput pattern often falls within the correct attractor basin. This is especially \nbeneficial in multistep tasks because cleanup can correct the inaccuracies introduced \nby the mapping network, preventing the system from gradually diverging from the \ndesired trajectory. \n\nProjection \n\n-\n= No projection \n\ntraining set size \n(% of all problems) \n\nFigure 4: Cleanup improves generalization performance. \n\nFigure 5 shows an example run of a trained network. There is one curve for each \nof the five result and carry pools, showing the degree of \"activity\" of the ultimate \ntarget pattern, t, for that pool as a function of time. Activity is defined to be \ne-lit-aIl2 where a is the current activity pattern and t is the target. The network \n\n\f/6 \n\nDonald Mathis, Michael C. Mozer \n\nsolves the problem by passing though the correct sequence of intermediate states, \neach of which are temporarily persistent. This resembles the sequence of conscious \nstates a person might experience while performing this task; each step of the problem \nis performed by an unconscious process, and the results of each of step appear in \nconscious awareness. \n\n1.0r----=====~------_\"\"::--_~\"\"-:-\"\"\"\"\"'----------, \n\nActivation of \ntarget pattern \nin each pool \n\n.8 \n\n.6 \n\n.' \n\n,:' :'\" \n: \n.' \" \n\" \" \n... :' \n,: ,'. \n\nresult digit I, \n- - carry digit I \n\nresult digit 2, \n----- carrydigit2 \n\n.......... result digit 3 \n\n- - - - - - - - - - ---\".:!/ \n~~~~IO~~~~2~O~--~3~O---~4~O---~50~--~ TIme \n\nFigure 5: Network solving the addition task in three steps \n\n3.2 CHOICE POINT SIMULATION \n\nIn many ordinary situations, people are required to make decisions, e.g., drive \nstraight through an intersection or turn left, order macaroni or a sandwich for \nlunch. At these choice points, any of the alternative actions are reasonable a priori. \nContextual information determines which action is correct, e.g., whether you are \ntrying to drive to work or to the supermarket. Conscious decision making often \noccurs at these choice points, except when the task is overlearned (Mandler, 1975). \n\nWe modeled a simple form of a choice point situation. We trained a module to \noutput sequences of states, e.g., ABCD or EFGH, where states were represented by \nunique activity patterns over a set of units. If the sequences shared no elements, \nthen presenting the first element of any sequence would be sufficient to regenerate \nthe sequence. But when sequences overlap, choice points are created. For example, \nwith the sequences ABCD and AEFG, state A can be followed by either B or E. \n\nWe show that cleanup allows the module to make a decision and complete one of \nthe two sequences. Figure 6 shows the operation of the module with and without \ncleanup following presentation of an A after training on the sequence pair ABCD and \nAEFG. There is one curve for each state, showing the activation of that state (defined \nas before), as a function of time. When the network is run with cleanup, although \nboth states Band E are initially partially activated, the cleanup process maps this \nill-formed state to state B, and the network then correctly completes the sequence \nABCD. Without cleanup, the initial activation of states Band E causes a blending of \nthe two sequences ABCD and AEFG and the state degenerates. 3 \n\nAlthough the arithmetic and choice point tasks seem simple in part because we \npredefined the set of well-formed states. However, because the architecture segre-\n\n3In this simulation, we are not modeling the role of context in helping to select one \nsequence or another; we are simply assuming that either sequence is valid in the current \ncontext. The nature of the model does not change when we consider context. Assuming \nthat the domain is not highly overlearned, the context will not strongly evoke one alter(cid:173)\nnative action or the other in the feedforward mapping, leading to partial activation of \nmultiple states, and the cleanup process will be needed to force a decision. \n\n\fOn the Computational Utility of Consciollsness \n\n17 \n\nActivity \nof state 0 .\u2022 \n\no. \n\n-\n\nstateB \n__ state C \n\n- - - -state D \n........ state E \n\nFigure 6: Decision-point task with and without cleanup \n\ngates knowledge of well-formed ness from knowledge of how to solve problems in the \ndomain, well-formedness could be learned simultaneously with, or prior to learn(cid:173)\ning the task. One could imagine training the cleanup network to autoassociate \nstates it observes in the domain before or during training using an unsupervised or \nself-supervised procedure. \n\n4 COMPUTATIONAL CONSEQUENCES OF \n\nPERSISTENT STATES \n\nIn a network of modules, a persistent well-formed state in one module exerts larger \ninfluences on the state of other modules than do transient or ill-formed states. As a \nresult the dynamics of the system tends to be dominated by well-formed persistent \nstates. We show this in a final simulation. \n\nThe network consisted of two modules, A and B, connected in a simple feedforward \ncascade. Each module's cleanup net was trained to have ten attractors, locally \nrepresented in a winner-take-all pool of units. The mapping network of module B \nwas trained to map the attractors of module A to attractors in B in a one-to-one \nfashion. Thus, state a1 in module A is mapped to {31 in module A, a2 to {32, etc. \n\nModule B was initialized to a well-formed state {31, and the output state of module \nA was varied, creating three conditions. In the persistent well-formed condition, \nmodule A was clamped in the well-formed state a2 for 50 time steps. In the transient \nwell-formed condition, module A was clamped in state a2 for only 30 time steps. And \nin the ill-formed condition, module A was clamped in an ill-formed state in which \ntwo states, a2 and a3, were both partially active. Figure 7 shows the subsequent \nactivation of state {32 in module B as a function of time. Module B undergoes a \ntransition from state {31 to state {32 only in the persistent well-formed condition. \nThis indicates that the conjunction of well-formedness and persistence is required \nto effect a transition from one state to another. \n\n5 CONCLUSIONS \n\nOur computational framework and simulation results suggest the following answers \nto our three key questions: \n\n\f18 \n\nActivity of \nstate ~2 in \nmodule B \n\n0' \n\nDonald Mathis, Michael C. Mozer \n\nwell-formed, persistent \nstate