20th WCP: Aristotle, Connectionism, and the Morally Excellent Brain

How can the brain be a mind, a conscious person? Recently, some philosophers have argued that human consciousness and cognitive activity, including even our moral cognition and behavior, can best be explained using a connectionist or neural network model of the brain (see Churchland 1995; Dennett 1991 and 1996). (1) Is this right? Can a mass of networked neurons produce moral human agents? I shall argue that it can; a brain can be morally excellent. A connectionist account of how the brain works can explain how a person might be morally excellent in Aristotle's sense of that term.

1. Connectionism

The brain receives input and somehow transforms it into output. How does it do it? In part because of the extraordinary technological feats achieved using digital processing computers, the brain has often been interpreted as a symbol manipulator and its cognitive activities as the transformation of symbols according to rules. By contrast, recent successes with parallel distributed processing computers have encouraged a connectionist theory of mind which regards the brain as a pattern recognizer and its cognitive activities as the transformation of neuronal activation patterns; however, these pattern transformations are not rule-governed processes, but straightforwardly causal processes in which networked units (neurons) excite and inhibit each other's activation level. The key to the connectionist theory is the neural network. The networks are layered such that one set of units receives the input stimulation and resets its pattern of activation levels; some set of in-between or "hidden" units receives as input the activation pattern of the first set of units; this set of hidden units in turn resets its activation pattern and stimulates the next level of hidden units; this process of activation pattern transformation continues until the set of output units resets itself accordingly and thereby sends the signal for whatever action is appropriate. The brain, then, is like a vector transformer exchanging one pattern of stimulation for another.

Churchland (1995) offers numerous examples of the brain's power to recognize patterns. Recognition of tastes, colors, and smells are all cases of the brain's receiving patterns of activation from stimulated sense organs, transforming those patterns by filtering them through layers of hidden units, and feeding those patterns to other neural nets to produce an appropriate output. Vectoral transformations at the level of hidden units are the key to categorization. Such transformations at any given level are produced as each cell's degree of activation is altered as a function of the stimulation it receives from each cell at the previous level together with its own adjustable synaptic connection "weight." It is the adjustable weights that make learning new categories possible, for reconfigured weights produce a new input-output function.

A connectionist system learns by adjusting its weights. Changing the weights changes the activation patterns produced at all levels, particularly the output level. Connectionist systems have been successfully trained in this way to recognize faces, words, and even mines on the ocean floor. In each case, the system's weights must be slightly readjusted numerous times to correct for output errors until it produces the appropriate input-output function. This method is called back-propagation as it uses a specialized computing rule (the generalized delta rule) to propagate the error measure that is calculated at the output units back through the network, thereby altering the weights. (2) Human brains probably don't work exactly like this, but they are chock full of recurrent pathways in the maze of neural connections which would make possible some analogous learning procedure. Like brains, connectionist systems trained in this way, not only recognize the right face or right word when they "see" it, they are also able to recognize new faces (that is, new presentations of old faces) or new words not in the original training set, and to recognize previously encountered faces or words whose inputs are significantly degraded.

In other words, connectionist systems can generalize or categorize on the basis of similarities. "There is a fairly clear sense," argue Bechtel and Abrahamsen, "in which connectionist networks are making similarity judgments: the similarity structure is implicit in the weight matrix. The weights are the means of treating similar inputs similarly." (3) And the weights, once set, determine stable patterns of activation within the local network. Perhaps concepts can be viewed as stable patterns of activation across an ensemble of units which determine further processing. (4) Indeed, analysis of the weight configurations and activation patterns of trained connectionist systems reveals that the multi-dimensioned space defined by stabilized middle layer activation patterns is partitioned into categories centered on prototypical cases; (the prototypes are defined by taking an average of the activation patterns produced by all relevant cases in the training set.) For example, in the facial recognition system the hidden layer activation patterns for male and female faces divide themselves along gender lines into two identifiable groups which cluster around prototypes for a male and a female face, respectively. Another connectionist system trained to read aloud from the printed page neatly, and as a natural result of the back-propagation training, separates vowels from consonants in its middle layer activation space. (5)

We may say, then, that the networks in both brains and artificial connectionist systems generalize concepts as they learn by reconfiguring their neuronal weights. And since

representations on hidden units result from the system's attempt to accommodate to its environment . . . , [this] learning procedure also gives the network a goal: maximizing the fit of its states to those of the environment (by minimizing error in producing outputs to inputs). Thus a teleological component is added. As a result, the representations developed in the hidden units subserve goals, and so can be thought of as representing information about things external to the system for the system. Hence, these representations are about the entities supplying the input. (6)

Bechtel and Abrahamsen here suggest that the activation patterns of the hidden units in a trained connectionist system are intentional states, for such representations are produced by the system in order for the system to pursue its ends.

We may conclude that since a connectionist system is capable of having concepts relevant to the achievement of its ends, ends the system is disposed to pursue under appropriate circumstances, then the brain, as connectionist system, has some of the abilities required of a system capable of moral excellence. For according to Aristotle, the morally excellent person must have some cognitive grasp of the end as well as a disposition to pursue that end. Aristotle also insists that moral excellence involves practical wisdom, the ability to recognize what needs to be done in order to achieve one's ends in various circumstances. As we shall see, connectionist systems are also capable of matching means to ends, of modifying their own behavioral output relative to the demands of a perceived situation. (7) In this paragraph I have introduced the main argument of my essay, which will be further developed in section three. However, I must first sketch some more details of Aristotle's ethical system.

2. Aristotelian Ethics

Aristotle argues that the good person is both practically wise and morally excellent. Moral excellence is a state of character in which one desires and chooses to pursue appropriate ethical ends or goals. Practical wisdom involves the rational capacity to deliberate well about what means are best for pursuing these ethical ends. Given experience and time, practical wisdom is teachable to the extent that practical reasoning skills are teachable. Moral excellence, however, is a matter of the formation of good habits; one must become disposed to pursuing ethical ends. This combination of moral character and practical intellect are essential to ethical action, for the former makes one do the end and the latter the means to the end.

The practically wise person excels at deliberating, calculating means to ends. But deliberation requires a grasp of the end as a starting-point, what Aristotle calls an arche or first principle; he claims that "in practical affairs that for the sake of which [i.e., the end] is a starting-point." (8) I have argued elsewhere that, according to Aristotle, the process through which one may acquire such ethical ends is inductive. (9) Good deliberation also requires an ability to see connections between means and ends, between the particular things-to-be-done and the goals to be pursued. Such means/ends connections are also starting-points of a second kind, starting-points for the induction of the end; "for the universals are from the particulars; of these it is necessary to have perception, and this is nous." (10) Nous, variously translated as "intuitive reason" or "comprehension" or "practical insight," is the capacity to perceive particular things-to-be-done as a means to an end. Thus, this ability to grasp connections between means and ends, the hallmark of practical wisdom, operates both in the process of induction to ethical ends and in the process of deliberation from ethical ends. (11)

So ethical ends are starting-points derived by induction, and means/ends connections are starting-points seen by the perceptual insight of nous. Where E is the end and M the means, here is how these two starting-points would function syllogistically as premises in deliberation:

1. E is what I should do. [ethical end starting-point]

2. M is (a means to) E. [means/end starting-point]

3. Therefore, M is what I should do. [conclusion]

This represents the logical form of deliberation, even though real agents need not think things through in this formal fashion. Likewise with the induction of the ethical end, which in form is the syllogistic reverse of the deliberation:

3. M1 (and M2, and M3, etc.) are what I should do.

2. M1 (and M2, and M3, etc.) are (means to) E.

1. Therefore, E is what I should do.

In this case, the M's in 3 are not necessarily recognized as means to an end; that recognition is represented by statement 2, the means/end starting-point. Statement 3 is a third kind of starting-point, acquired by habit. Aristotle claims that moral excellence or virtue comes about as a result of habit, that we become just by doing just acts; this is why the habits one forms from training in youth make "a very great difference, or rather all the difference." (12) Thus the M's in 3 indicate the cases in the training set, for example the various commands that adults issue to children learning to be just. ("Give Bobby a chance to play," "Don't hit her, she didn't do anything to hurt you," "You have to give his ball back, it's not yours," "You broke the window, now how do you expect to pay for it?") Although we sometimes tell children why they must behave in certain ways, we often as a starting-point simply tell them that they must so behave and then leave it to them to induce the ends as they accumulate more experience. (13)

Along the way to moral excellence and practical wisdom, then, Aristotle recognizes three kinds of starting-points, ethical ends, means/ends connections, and particular things to be done. These three kinds are acquired, respectively, by induction, by the perceptual insight of nous, and by habit. (14) By induction one knows the end, by habit one desires (pursues) the end, and by perceptual insight (nous) one recognizes how to reach the end. If a person is to achieve moral excellence s/he must have these cognitive capacities, as must any connectionist system capable of moral excellence, like a human brain for instance.

3. The Psychological/Cognitive Capacities of the Morally Excellent Brain

Aristotle's account of the psychological capacities of the morally excellent person is compatible with a connectionist theory of the mind/brain. The good person's brain must be able: to form morally relevant concepts (by induction) to recognize ethical ends; to develop appropriate behavioral dispositions (by habit) to pursue ethical ends; and to learn practical reasoning skills (by nous) to perceive the means to the ends. This collection of the brain's cognitive capacities are best accounted for by a connectionist theory of the mind/brain.

We have already seen that connectionist systems have the ability to generalize categories and then apply the concepts to new or modified cases. By reconfiguring its neuronal weights, my brain can learn to recognize the faces of more than fifty new students in just a few weeks time; and then, as a matter of learned habit, my brain can produce the right name for the right face in a variety of circumstances. Stable activation patterns have been established in the hidden units of my brain's neural nets. And this same training that establishes the habits establishes the concepts. When the "Kari Middleton" category is activated I am disposed to produce her name; and given the complexity of the connectionist system in my head, I am also disposed to behave in her presence in ways bound up with my categories of "female" and "student." If connectionism provides a plausible account of this brain training, and it does, then it may also account for my tendency to respect my students as persons and to treat them fairly. Parents and teachers insisted that I share my toys, that I listen while others are talking, that I not deceive others to gain an advantage, etc. From this training set, Aristotle would argue, I have generalized a concept of fairness or justice, and have become disposed by habit to pursue the goal of behaving fairly and justly. Connectionism provides a likely account of the brain's role in this Aristotelian process. The morally excellent brain recognizes (by induction) and pursues (by habit) ethical ends. Given the proper stimulation, learned vector patterns are activated in the hidden units of the brain, and this activation is at once both a recognition of the goal and a disposition to pursue it.

Aristotle also insists that moral excellence involves practical wisdom, the ability to recognize (by nous) what needs to be done in order to achieve one's ethical ends. This ability requires the identification of both an appropriate categorization of the perceived situation and an appropriate action to be taken to meet the demands of the situation. As we have just seen, for a connectionist system to categorize a situation, even a new one, is to dispose itself to pursue a certain goal and to produce a certain output. In the case of a relatively simple single-task connectionist system like the face-recognizer, this may well be the end of the story. Its end goal is to identify faces; categorizing a face as Bill's faces already disposes it to pursue this end by the means of identifying the presently perceived face as Bill's face; but since the particular manner of its output (words on a screen perhaps) is presumably hardwired and arbitrary, honing its means to any further detail is beyond its ken. However, for a system requiring motor skills for interaction with the environment, the story may continue. Initially categorizing the situation might not be sufficient for triggering the appropriate means for achieving the goal.

Suppose you are a crab in search of a tidbit. You categorize your present situation as "food for me off to the right," that is, you are stimulated into a neural activation pattern that disposes you toward the goal of grabbing and ingesting the food. But you aren't hardwired to display your categorization on a monitor screen, nor would you find that output fulfilling. You need to actually grab the food and get it in your mouth before it moves away. Thus you need a well-timed transformation of sensory vectors into motor vectors, a transformation likely to involve a variety of interacting neural nets generating complex sequences of activation vectors. (15) Only in this way may your goal be achieved by moving your arms and pincers in the right way at the right time. (16) This sort of matching of means to ends, this sensorimotor coordination, is, according to Paul Churchland, "where intelligence begins: in the brain's capacity for executing principled sensorimotor transforms. . . . [and] this know-how is embodied in the personal configuration of the brain's synaptic weights." (17) Aristotle would call such intelligence or know-how, when applied in the realm of ethics, practical wisdom, the ability to recognize and do what needs to be done in order to achieve one's ends, in the right manner at the right time.

You are no longer a crab in search of a tidbit, but an undergraduate in search of lunch. Waiting in line at the lunch counter, you categorize your present situation as "ham sandwich for me off to the right." You are stimulated into a neural activation pattern that disposes you toward the goal of grabbing and ingesting the sandwich. But you have been reading about animal rights issues in your ethics course. You hesitate as another category is activated, "suffering person nearby." Activation of this category normally disposes you toward alleviating the suffering of others. Pigs aren't persons, but they are similar in some ways. "Suffering pigs" begins to occupy a point in vector space uncomfortably near "suffering persons." You find yourself stimulated into a neural activation pattern that disposes you to not grab the sandwich. To thus extend a category or prototype in one's moral deliberations is to insightfully perceive another means to one's ethical ends. (18) The undergraduate's imaginative connection between suffering pigs and suffering persons is, among other things, the perceptive insight that certain particular acts are means to previously established ends; refusing to eat pigs is a means to alleviating suffering. This insight into means/end connections is the work of nous, to use Aristotle's term; and nous is the mark of practical wisdom. (19)

Thus the connectionist brain does have the cognitive capacities to achieve moral excellence and practical wisdom, in Aristotle's sense of those terms. It can know the right end by induction, desire the right end by habit, and perceive the right means by nous. In the end, what makes a brain morally excellent and practically wise is the proper setting of its neuronal weights to establish hidden layer activation patterns capable of transforming similar inputs into similar outputs. That may sound too mechanical. But by filling in the mechanical details, one sees that some of the capacities of the connectionist brain machinery are just those capacities required for the production of ethical behavior. Thus the human condition is both material and moral; we are brain-controlled bodies with ethical values.

Notes

(1) Whereas Paul Churchland insists on accounting for the cognitive capacities and even consciousness of the human brain by appealing directly to the amazing features of connectionist systems, Daniel Dennett argues that "even if we succeeded in explaining [complex cognitive] processes at the level of synapses or bundles of neurons, we would be mystified about other aspects of what must be happening" (1991, p.193). For example, what is unique about human cognition and consciousness is our ability to store and then manipulate concepts or ideas as needed in a serial fashion; hence human brains must have the capacity to re-represent the representations that occur in their neural networks. Thus Dennett argues that "[c]onscious human minds are more-or-less serial virtual machines implemented . . . on the parallel [connectionist] hardware that evolution has provided for us" (1991, p.218). This virtual machinery involves the "improvements we install in our brains when we learn our languages, [which] permit us to review, recall, rehearse, redesign our own activities, turning our brains into echo chambers of sorts, in which otherwise evanescent processes can hang around and become objects in their own right. . . . [W]ithout a natural language [a creature] has no talent for wresting concepts from their connectionist nests and manipulating them" (1996, pp.155, 159). Churchland responds that Dennett's virtual serial machine is an unnecessary level of explanation: "The fact is, there exists a different way, a much more natural and effective way, of accounting for the well-behaved temporal unfolding of consciousness, and it has nothing essential to do with serial computers or language-like processing. The alternative lies in the dynamical behavior of real (not virtual) recurrent networks, with their dramatic ability to generate complex representations with a continuously unfolding temporal dimension" (p.267).

Both Churchland and Dennett agree that the hardware of the brain is connectionist. Dennett believes that a virtual meta-machine must be running on top of that connectionist hardware to account for the sophisticated cognitive activity of conscious humans; Churchland believes that features of the connectionist system alone are sufficient. Is Dennett positing an unnecessary "higher" level of explanation? Or is he developing useful ways of describing the intense complexity of the connectionist system that is the human brain? I will be following Churchland's lead in this essay, although Dennett's alternative would provide an important contrast.

(2) For a sophisticated description of this learning procedure see chapter three of Bechtel and Abrahamsen 1991. For a much simpler version see Churchland 1995, pp.42-45.

(3) Bechtel and Abrahamsen 1991, p.121.

(4) Bechtel and Abrahamsen 1991, p.138.

(5) See Churchland 1995, chs. 3-4.

(6) Bechtel and Abrahamsen 1991, p.128.

(7) Matching means to ends is in part a problem of sensorimotor coordination. According to Churchland (1995), networked systems can solve this problem.

The problem is how to produce behavior that is appropriate or intelligent relative to a perceived situation. . . . If the external environment is represented in the brain with high-dimensional coding vectors; and if the brain's "intended" bodily behavior is represented in its motor nerves with high-dimensional coding vectors; then what intelligence requires is some appropriate or well-tuned transformation of sensory vectors into motor vectors! What sort of mechanism might perform such a task? We already know the answer: a multi-layered neural network, with a well-configured matrix of synaptic connection weights. (p.93)

(8) Nicomachean Ethics 1151a16. Translations are mine.

(9) DeMoss 1990, "Acquiring Ethical Ends." My account of Aristotelian ethics in the present essay is supported by arguments presented in DeMoss 1990 which involve detailed analysis of passages from Nicomachean Ethics.

(10) Nicomachean Ethics 1143b4-5.

(11) Aristotle warns, "Let it not escape our notice that the arguments from the starting-points and those to the staring-points differ" (Nicomachean Ethics 1095a30-32).

(12) Nicomachean Ethics 1103b24-25.

(13) Aristotle often stresses the necessity of training in good habits as a foundation for moral growth; he writes, "Wherefore the one who will be listening competently to what concerns good and just things and in general the science of politics must have been well-trained in his habits. For the starting-point is the that, and if this should be sufficiently apparent, one will not need in addition the reason why" (Nicomachean Ethics 1095b6-8). (See also 1098a33-b3.)

(14) It is not surprising, then, that Aristotle claims that "some [starting-points] are seen by induction, some by perception, and some by a certain acquisition by habit, and others also in other ways" (Nicomachean Ethics 1098b3-4).

(15) See Churchland 1995, chapter 5, "Recurrent Networks: The Conquest of Time."

(16) One is reminded here of Aristotle's warning that there are many ways to go wrong in the choice of means. One must do the right thing in the right way at the right time. (See Nicomachean Ethics 1142b16-28.)

(17) Churchland 1995, p.95.

(18) In his book, Moral Imagination: Implications of Cognitive Science for Ethics (1993), Mark Johnson explains,

A central part of our moral development will be the imaginative use of particular prototypes in constructing our lives. . . . Many of our moral problems stem from questions about permissible metaphorical extensions from the prototype to noncentral members . . . . Our moral deliberations will be about whether, for instance, certain 'higher' mammals ought to be understood metaphorically as persons, and therefore accorded certain rights. (pp.192 and 195)

(19) See DeMoss 1990, pp.72-73, where I discuss the nature of practical reasoning and argue that the "recognition that certain particular acts are means to an end . . . requires an imaginative leap, a special kind of perception, nous."

References

Aristotle, Ethica Nicomachea, ed. Bywater (1894) (Oxford: Oxford University Press).

Bechtel, W. & Abrahamsen, A. (1991), Connectionism and the Mind: An Introduction to Parallel Processing in Networks (Cambridge, MA: Basil Blackwell).

Churchland, Paul (1995), The Engine of Reason, the Seat of the Soul: A Philosophical Journey into the Brain (Cambridge, MA: MIT Press).

Churchland, Paul (1988), Matter and Consciousness: A Contemporary Introduction to the Philosophy of Mind, revised edition (Cambridge, MA: MIT Press).

DeMoss, David (1990), "Acquiring Ethical Ends," Ancient Philosophy, 10, pp. 63-79.

Dennett, Daniel (1991) Consciousness Explained (Boston, MA: Little Brown).

Johnson, Mark (1993) Moral Imagination: Implications of Cognitive Science for Ethics (Chicago: University of Chicago Press).