anarkhos

Introduction

In this paper we will consider three Heideggerian objections to Symbolic Artificial Intelligence (AI). We will attempt to show that Connectionist AI answers these objections and propose that some of Heidegger's conceptions of the problem of mental representations be rethought in light of the Connectionist model.

Hubert Dreyfus, in his well known role as nemesis of the field of Symbolic AI, attacked that line of research primarily on Heideggerian grounds. Pointing out that objects, for humans, are not context free entities, but equipment, embedded in a cultural world, in which they are enmeshed in a network of relations with other equipment, he concluded that digital computers, supporting symbolic representations with intentional content, could not reason intelligently in this human world.

The human world, then, is prestructured in terms of human purposes and concerns in such a way that what counts as an object or is significant about an object already is a function of, or embodies, that concern. This cannot be matched by a computer, which can deal only with universally defined, i.e., context-free, objects.... In Being and Time Heidegger gives a description of the human world in which man is at home, on the model of a constellation of implements (Zeuge), each referring to each other, to the whole workshop and ultimately to human purposes and goals.¹

We will refer to this objection to traditional AI as Objection One. Objection One may be restated: In order for a machine's cognition to count as intelligent its conception of the world must be in terms of its purposes and concerns, and the elements of that world must be considered primarily in terms of their equipmental character relative to these purposes and concerns.

Traditional AI relies upon the validity of the philosophical tradition, beginning with Plato, which contends that only by removing the facts and relations of the world from our experience and considering them from this objective standpoint can we engage in intelligent behavior and understand the world in which we find ourselves. Thus, symbolic approaches to AI have sought to couch learning and problem solving in terms of theories, embodied in rules, expressing knowledge in terms of symbolic reference to real world objects and states.

Heidegger would object that it is in worldly experience that objects are, and thus it is with this experiential knowledge that we are actually concerned.

The kind of dealing which is closest to us is as we have shown, not a bare perceptual cognition, but rather that kind of concern which manipulates things and puts them to use; and this has its own kind of 'knowledge'.²

Similarly, Dreyfus contends:

When we try to find the ultimate context-free, purpose-free elements, as we must if we are going to find the ultimate bits to feed a machine - bits that will be relevant to all possible tasks because chosen for none - we are in effect trying to free the facts in our experience of just that pragmatic organization which makes it possible to use them flexibly in coping with everyday problems.³

Even if a machine could operate in the world with some degree of success using a non-experience laden theory, the theory, in its impoverishment of experience would not be able to capture all relevant meaning.

Of course, Heidegger would not claim that his derivation of the predicate calculus shows that one cannot represent everyday intelligibility in this or some other calculus. What he would claim is that once one sees that symbolic logic is the result of the progressive decontextualization and impoverishment of our everyday language for pointing out equipment and its aspects, one will no longer be inclined to believe that logic, although a universal and unambiguous medium, is an appropriate form in which to express all meaning.⁴

For our purposes in this paper we will refer to this objection to traditional AI as Objection Two. Objection Two may be restated: In order for a machine to behave intelligently, its knowledge must be experience laden.

Finally, we will consider the following objection. When humans view a partial or ambiguous scene, hear a sentence fragment or slightly garbled utterance, or hear the first few notes of a familiar melody, they quickly fill in the absent material, as if the relevant knowledge was somehow pulled in to fill the vacuum. Moreover, how frequently does it happen that an experience pulls in a world of relevant knowledge and previous experience so as to enlighten our understanding of that experience! As Dreyfus remarks:

When we perceive an object we are aware that it has more aspects than we are at the moment considering. Moreover, once we have experienced these further aspects, they will be experienced as copresent, as covered up by what is directly presented. Thus, in ordinary situations, we say we perceive the whole object, even its hidden aspects, because the concealed aspects directly affect our perception.... A machine with no equivalent of an inner horizon would have to process this information in the reverse order: from details to the whole. Given any aspect of an object, the machine would either pick it up on its receptors or it would not. All additional information about other aspects of the object would have to be explicitly stored in memory - in Minsky's sort of model - or counted out again when it was needed. This lack of horizons is the essential difference between an image in a movie or on a TV screen and the same scene as experienced by a human being.⁵

Heidegger, writing of the same phenomenon states:

A sign is not a Thing which stands to another Thing in the relationship of indicating; it is rather an item of equipment which explicitly raises a totality of equipment into out circumspection so that together with it the worldly character of the ready-to-hand announces itself.⁶

We will refer to this objection as Objection Three. Objection Three may be restated thus: An intelligent machine must be able to bring to the fore all knowledge relevant to the element of its perception with respect to its context and purposes.

We will now attempt to show how the Connectionist model of cognition answers these objections.

Objection One

Most implementations of systems incorporating the Connectionist model rely upon the input of a vector of features extracted by some sort of analog to digital converter. This input vector is the neural network's sensory input. Unlike traditional computer systems, neural network systems are not programmed. They are trained on sets of input vectors. Depending upon the training method selected, the training may occur with or without the assistance of a teacher. In either case, neural network system learn by experience. Leaning consists of carving up an activation vector space into regions (prototypes) into which input feature vectors are transformed. The network itself carves out these prototypes according to learning rules which are used to modify the weights connecting the units which comprise the network. Once trained, classification of an input pattern is merely a matter of testing the transformation of an input vector to determine which prototype it maps into (or which prototype it maps closest to).

If we were to take this model of representation to its logical conclusion and imagine that we were to train a robot to become a being capable of behaving intelligently within our world, we would train the robot much the same way as we train children. For example, if we wished to train a child to recognized cups, we would not arrange an endless array of cups on a conveyor belt and ask the child, as each cup passed by, "What is that?", affirming or denying the child's response and then making the proper identification. Rather, we would, in the course of every day activity, show the child a cup whenever it became part of her activity and then ask the child what it was, explain that it was a cup and that cups are used for drinking. "They hold liquid. They are used in just such a way."

Heidegger, in the following passage, confronts this equipmental nature of our intentionality:

Only because equipment has this 'Being-in-itself' and does not merely occur, is it manipulable in the broadest sense and at our disposal. No matter how sharply we just look at the 'outward appearance' of Things in whatever form this takes, we cannot discover anything ready-to-hand. If we look at Things just 'theoretically', we can get along without understanding readiness-to-hand. But when we deal with them by using them an manipulating them, this activity is not a blind one; it has its own kind of sight, by which our manipulation is guided and from which it acquires its specific Thingly character.⁷

Churchland, in his defense of connectionism against an antinaturalist argument, contends:

It therefore should not be a surprise that a human infant comes to recognize and respond to cultural features that resist definition in terms of notions like mass, charge, length, and so forth, because the most dominant "teacher" in the local environment is the culture into which the infant is born. The set of weights that constitutes a child's developing consciousness is continually being shaped by the linguistic, conceptual, and social surround. The developing brain comes to reflect the elements and structure of that surround in great detail. For that is what networks do. What shapes them is the stimuli they typically receive, and the subsequent corrections in their responses to which they are typically subject. Small wonder we become attuned to the categories of the culture that raises us.⁸

In fact, for most of us, Euclidean Geometry is probably our first experience of learning without reference to utility. In virtual all of our learning experiences before that point, we are taught to identify things in terms of their use, within our culture. We learn the prototypical geometric configuration of a telephone after we experience many of them in use. It is by using them and watching other's use them that we discover what a telephone is. Seldom are we surprised when encountering an oddly shaped telephone. However, we are quite disturbed when a telephone ceases to function as it should. When we use a telephone, its physical configuration drops out from our consciousness and it becomes an extension of our purpose. It is this purpose which defines the telephone essentially, and it is likely those features which characterize its purpose that we use primarily to distinguish telephones from non-telephones.

The peculiarity of what is proximally ready-to-hand is that, in its readiness-to-hand, it must, as it were, withdraw in order to be ready-to-hand quite authentically. That with which our everyday dealings proximally dwell is not the tools themselves. On the contrary, that with which we concern ourselves primarily is the work - that which is to be produced at the time; and this is accordingly ready-to-hand too.⁹

Similarly, our robot would learn to identify objects in terms of their purposes. Just what input vectors would the robot use? From its entire repertoire of sensory input: vision, sound, touch, heat, taste, bodily position, and their temporal orderings, it would learn just which sensory combinations were necessary for cup identification by using cups. It would construct its own input vectors of the relevant combinations and create an appropriate feature space, mapped to the appropriate activation vector space, which, in time, would be carved up into a useful set of prototypes. These prototypes would embody the utility and purpose of the cup with respect to the robot. Thus for a complete and autonomous connectionist being, the being would conceive of the world in terms of its purposes and concerns. The objects of just such a being's world would be considered in terms of their equipmental character. Thus have we addressed Objection One.

Objection Two

In Connectionist AI, it is the pattern of activations within a network which embodies intentionality. Contrasting this with the utilization of symbols in traditional AI to refer to objects in the real world, we can see that Connectionist AI does not separate experience from representation as does the traditional symbolic representation. The pattern of activation within a trained neural network, is the causal consequence of the impression upon the network of a pattern from sensory space and the prior experiential tuning of the network. The network's configuration of weights is the embodiment of experience. The representation impressed upon the network from an instance is the causal consequence of that instance. In this sense, there is a unique mapping of instance to mental representation. The instance's mapping onto a prototype gives the instance 'meaning', but, reciprocally, the prototype itself may be impacted by the instance. Thus, the knowledge embodied in a neural network is experience laden. Experience is prior to knowledge, but understanding is contingent upon knowledge.

Interesting phenomena result from conditions of "error" within this representation:

This also gives a perspective on the profile of perceptual misjudgments, such as mistaking poison oak for Oregon grape, or perceptual illusions, such as seeing a letter where one has been omitted (grapfruit), or omitting a word where there there is an extra. It gives a general means of understanding prototype phenomena in perception and cognition - for example, why carrots are rated as more typical vegetables than is corn, and why corn is rated as more typically a vegetable than is parsley.¹⁰

Consider, for example, what would happen if a trained neural network were impressed upon by the feature vector of a pattern for which no category had yet been learned. Those categories which the instance most closely resembles, with respect to those features discriminated as salient by the network, would be activated and thus a best guess would be made. As a corollary of this phenomenon, such a network can represent instances which do not exist. Just as in our imaginations, these combinations of features would share characteristics of various other categories and thus provide reference to imaginary entities, giving them 'meaning'.

Knowledge in Connectionist AI is clearly not context free. As it is also experience laden and embodies meaning in experience, the concern of Objection Two has thus been met.

One important difference between the connectionist and traditional symbolic views of the interface with the world is that in the connectionist account, representations will not be arbitrary.... That is, two-layer pattern recognition networks, using a learning rule such as the delta rule, modify themselves toward weights that directly reflect environmentally-given relations between input and output patterns. Multilayered networks additionally determine what higher-order information (features or microfeatures) should be encoded in hidden units.¹¹

Objection Three

An AI system employing the traditional symbolic representation is limited in its ability to bringing to the fore relevant features given a set of sensory input. All pattern matching in this paradigm requires some sort of search. Searches take time and it is not always clear where they should be terminated. Furthermore, to call in additional information when relevant features are not immediately present may require an explosive branching of further references denoted by the symbols currently brought to the fore and their relations. As we shall see, connectionist representations are quite adept at these tasks.

Connectionist representations are learned by experience. The association of relevant features goes part and parcel with the prototype representation. Additionally, neural networks do very well at vector completion. Utilizing vector completion, most neural networks can fill in the relevant values missing from an input vector and classify the input:

In addition to recognizing patterns, they can also complete patterns by filling in what was not present in the input. This capacity is a general feature of connectionist networks.¹²

This amounts to being able to see the big picture from partial information, which is very much an answer to Objection Three.

In principle, vector completion allows for filling in of relevant visual information even when the input vector is basically auditory. Thus once the Daisy representation is accessed, a normal subject may be able to go on and describe Daisy's visual properties in loving detail.¹³

Which may go a long way in a being's purposeful deliberation of the entity's (equipment) within its world:

Such a representation allows the creature to anticipate the aspects of the case so far unperceived, and to deploy practical techniques appropriate to the case at hand.¹⁴

Thus, the solution to Objection Three is embedded in the representation employed by the connectionist model.

Connectionist Implications for Heidegger and Dreyfus

In this section we will consider what, if any, implications the connectionist model might have on Heideggerian thought. We will consider two issues: the embodiment of Dasein's being-in-the-world as mental states and the Heideggerian concept of Falling as plasticity.

Heidegger conceives of Dasein's basic mode of being as defined in terms of the cultural relations and equipment in which Dasein finds itself imbedded. Being-in-the-world is Heidegger's term for this mode of being. Being-in-the-world is prior to the mental. The mental is represented in terms of this being-in-the-world. This conception is parallel to, but distinct from, the connectionist representation.

Dreyfus, in explaining Heidegger's conception of Dasein's being, writes:

Since Descartes, philosophers have been struck with the epistemological problem of explaining how the ideas in our mind can be true of the external world. Heidegger shows that this subject/object epistemology presupposes a background of everyday practices into which we are socialized but that we do not represent in our minds.¹⁵

Heidegger questions both the possibility and the desirability of making our everyday understanding totally explicitly. He introduces the idea that the shared everyday skills, discriminations, and practices into which we are socialized provide the conditions necessary for people to pick out object, to understand themselves as subjects, and, generally, to make sense of the world and of their lives. He then argues that these practices can function only if they remain in the background.¹⁶

Cognitivism, or the information-processing model of the mind, is the latest and strongest version of formal representations and thus seeks to explain human activity in terms of a complex combination of logically independent symbols representing elements, attributes, or primitives in the world. This approach underlies decision analysis, transformational grammar, functional anthropology, and cognitive psychology, as well as the belief in the possibility of programming digital computers to exhibit intelligence. Heidegger's view on the nonrepresentable and nonformalizable nature of being-in-the-world doubly calls into question this computer model of the mind.¹⁷

The connectionist representation supports the cultural imbeddedness of equipment by allowing objects to be represented in terms not only of primary sensory input, but in terms of derivative relations of cultural imbeddedness. In fact, as we have seen in the method by which children learn and its parallel implications for connectionist systems which might be similarly trained, the features of objects presented for training would naturally be presented in terms of these cultural and utilitarian relations.

Unlike Heidegger, however, the connectionist representation does indeed provide for mental representations of Dasein's cultural imbeddedness in the form of activation spaces whose dimensions include features which denote learned cultural and purposeful relations. Similar to Heidegger's model, these features would need not be conscious. Thus, phenomenologically speaking, the connectionist model, with respect to this aspect, would be isomorphic with Heidegger's being-in-the-world.

The connectionist model is predicated on the concept of plasticity. Plasticity is the ability to restructure representations on the basis of experience. In this sense, connectionist systems are self created. If the human mind is a connectionist system, then much of its being as a cognitive entity is self created and thus is its nature not completely defined by its physical structure. Phenomenologically, however, the connectionist 'mind' cannot directly access its internal representations by conscious introspection. Having learned to interpret its behavior in terms of culturally acquired prototypes embodying the theory of folk-psychology, it is not aware that its nature is to a large degree self defined. Thus it takes its behavior as a manifestation of immutable nature. This results is a misinterpretation of self. As experience changes the nature of the 'mind' it is confronted with its self created nature which leads to crisis. The implication here is that it is possible to formulate Heidegger's concept of Falling in connectionist terms.

Conclusion

Dreyfus' criticism of Symbolic AI, on Heideggerian grounds, has appropriately led many in the field of Artificial Intelligence to rethink their formal approach to the representation of knowledge. The emerging Connectionist model has provided an optimistic alternative for those attempting to propose a materialist model of mind. It addresses the problem of the experience laden nature of intentionality, the cultural imbeddedness and purposefulness of the phenomenon we encounter in our everyday existence, the associative and interpretive nature of our understanding, and the self-createdness of our nature.