MCB

IX. Internat. Kant-Kongress, Berlin, 26.3 - 31.3.2000 - Section 18 - Kant und die Folgen

Learning from Kant How To Construct Robots
Suggestions for an Engineering Approach to the Analytic of Concepts

Marco C. Bettoni

My goal in this paper is to sketch a new approach to Kant's Critique of Pure Reason (Critique), one specifically adapted to the needs of (cognitive) scientists and engineers engaged in the design and construction of autonomous (learning) robots.   The ideas that I'm going to present are based on my previous work (M. Bettoni, 1988-1999[1]) which consists in some experiments in linking Kant's 'Analytic of Concepts' with 'Knowledge Engineering' on the basis of a cybernetic approach.

Because classical AI is facing a number of fundamental problems, in the last decade (cognitive) scientists and engineers involved in the development of robots on the basis of a multidisciplinary study of mind, brain and behavior (cognitive science) have put their efforts in developing new, alternative models of cognition, like Connectionism, "New AI" and humanoid (subsumption) robots. Surprisingly, also these new approaches are dominated either by rationalists or by empiricists positions[2].

One dominant rationalist position is strict nativism: it implies that cognition is crucially dependent on the a priori (in the sense of "pre-specified") definiton of its world model. This is unsatisfactory in robot design, because such an "innate world model" implies a behavior in which the robot uses knowledge specified independently from its experiences. It therefore can only undertake stereotypical behaviour as defined by the knowledge base supplied by a programmer.

Rodney Brooks has proposed subsumption architecture (COG project), an opposite approach to nativism, which relies on reactions determined by the sensed states of the external world and considers a world model as unnecessary[3]. What is missing in such a purely "empiricist" robot is the ability to introduce "order and regularity" (A 125) of a higher level into its impressions ("Eindrücke") and sensations. In Kant's terms, the subsumption architecture lacks "the concepts which render possible the formal unity of experience, and therewith validity (truth) of empirical knowledge" (A 124). Thus it is not astonishing, that pure subsumption robots have not yet demonstrated the ability to perform complex real tasks.

A new approach, which attempts to overcome these problems by integrating (like Kant did more than 200 years ago) the viable aspects of the rationalist and the empiricist positions, has been recently proposed by Paul Verschure at the University of Zürich [2, p. 622] who argues that "before we can address the issue how a system uses its knowledge the question of how this knowledge is acquired and retained needs to be explored." This means, that the robot's world models "cannot be fully pre-specified but need to be acquired" [2, p. 622] through its interactions with the world [4]. This raises the question of how experience (learning) can be explained. Unfortunately Verschure by programmatically excluding a priori knowledge excludes not only innate, pre-specified knowledge but also experience-independent elements, which could play a major role in answering the question of experience.

In this paper we address this question making the assumption that concepts and objects play a fundamental role in experience and we claim that reading Kant's Analytic of Concepts in a way adapted to the needs of robot engineering could contribute important ideas for specifying the role that a priori knowledge could play in the mechanisms of a robot that is able to acquire its own world model through interaction with its environment.

In order to develop such a robot-engineering perspective on Kant’s work on the mind let us first consider its most general need: every conception and design of a robot's cognitive faculties needs - in order to be implemented as a physical machine - to be expressed in terms of functions and operations, otherwise no implementation would be possible. Silvio Ceccato [5], a pioneer of the cybernetic approach to cognition, wrote nearly 35 years ago [5, p.5]: "The activity of the mind must enter the machine exactly as activity and completely as activity, so that what natural organs perform in a human being is performed by the artificial organs of the machine". As a consequence, understanding Kant's epistemology from a robot engineering perspective means that the scientist or engineer should look for "functional" and "operational" descriptions in his work. The general principle of such an engineering approach is then a very simple cybernetic principle:

1. operational descriptions of cognitive faculties, by distinguishing operators, operations and operands
2. operational explanations of cognitive functions, by distinguishing the mechanisms implemented in the faculties.

To show that an operational approach is justified (by Kant himself) let us consider the two major concerns in Kant’s work, the metaphysical and the functional. Kant mentions these two concerns and their relation very early in the evolution of the Critique. In a letter to Marcus Herz (21.2.1772) Kant writes about his search for "die Quellen der intellektuellen Erkenntnis" which he considers to be necessary for determining "die Natur und Grenzen der Methaphysik". From this way of linking cognition and metaphysics we can hypothesize Kant's early conviction that the metaphysical goal of the Critique (to answer the question "What can we know?" by justifying our right to make synthetic a priori knowledge claims) requires the pursuit of a functional goal which consists in answering the question "How do we know?" by setting up a functional model of mind - the source of our intellectual knowledge - as a coherent, systematic model (an architecture) of mental faculties, processes, operations and products like in Kant's conception of "ein Gliederbau worin alles Organ ist" [B XXXVII].

Nine years later, in the preface to the first edition of the Critique, Kant points out that his "Deduction of the pure concepts of understanding" (the heart of the Critique) has two sides (A XVI-XVII): "The one refers to the objects of pure understanding ... The other seeks to investigate the pure understanding itself, its possibility and the cognitive faculties upon which it rests;".

As Kitcher observes "Kant was concerned with both the possibility of vindicating synthetic a priori knowledge and with the possibility of cognition itself" [6, p.6]. These two concerns specify two sides of the Critique, a metaphysical inquiry into the nature of thought and the limits of knowledge and a functional doctrine of cognitive faculties united like in an "organised body" (B XXIII), two sides that Kant presents as tightly intertwined especially in the Aesthetic, Analytic and Dialectic.

Disentangling these two sides is fundamental to the specification of a cognitive approach to Kant: we must gain a clear view of what belongs to the metaphysical side and what to the functional doctrine. In the operational perspective proposed here this task becomes relatively easy, because the principle of 'Operational Interpretation' gives a clear criterion for distinguishing the functional doctrine needed by robot engineering. In that principle we use the term "operational explanation" that we have derived as follows from what Maturana [7] calls a "scientific explanation" in his analysis of objectivity:

That is, operational explanations are propositions of generative mechanisms.

Given our principle of operational interpretation and our definition of operational explanation, our robot engineering approach can be further specified by the following central questions:

1. What is, from that operational point of view, the fundamental question that Kant answers in the Critique ?
2. What is, from that operational point of view, the global answer to that question?
3. Where, from that operational point of view, does Kant develop his answer to that question?
4. How, from that operational point of view, does Kant answer that question ?

The first two questions can be seen in their relevance to robot construction if one considers that everything that a robot interacting with the world does, can be subsumed under the general category of relating cognitive states on one side with objects of its environment on the other side. In fact, by constructing a robot we are going to introduce into our scientific considerations and experiments a kind of machine completely different from all other machines, namely a kind of "artificial observer" structurally coupled to the environment.

As a consequence, in designing the mechanisms that will allow such an artificial observer to interact with the world a scientist or engineer needs to answer the fundamental question of the relation between knowing and reality. Also it should be noted that - regardless of whether we are aware of it or not - any design of a learning robot interacting with the world entails an answer to that fundamental question. For these reasons, another pillar of our engineering approach to Kant consists in focusing on Kant's contribution to that fundamental question. In the previously quoted famous letter to his friend Markus Herz (21.2.1772) Kant wrote that he was planning to publish a work in which he wanted to answer the question "auf welchem Grunde beruhet die Beziehung desjenigen, was man in uns Vorstellung nennt, auf den Gegenstand" (on which ground is founded the relation between what we call 'Vorstellung' (mental construction) in us and the object). Fifteen years later, in the Critique of pure reason, Kant presents the answer to that question as the central hypothesis of his work by saying that objects must conform to our knowing [8]. He also emphasizes that this "revolution" constitutes the main purpose of the Critique (BXXII).

In trying to understand Kant's "Copernican Revolution" (objects must conform to our knowing) from a robot-engineering perspective, the real challenge lies in a critical reflection upon the specific components of Kant's hypothesis, which amounts at asking 3 crucial questions:

What is, in operational terms

1. Kant's conception of "object" ?
2. Kant's conception of "must conform to" ?
3. Kant's conception of "(faculty of) knowing" ?

The basis for answering these questions can be found in the Analytic of Concepts (B90 to B 169). Kant himself draws the reader’s attention to what he considers the central part of his argumentation as follows: "I know no enquires which are more important for exploring the faculty which we entitle understanding, and for determining the rules and limits of its employment, than those which I have instituted in the second chapter of the Transcendental Analytic under the title Deduction of the Pure Concepts of Understanding. They are also those which have cost me the greatest labour ..." (A XVI).

As an example, consider, in the Deduction, Kant's summarized explanation of the Copernican hypothesis in the crucial § 17 in the B edition of the Critique: "Verstand ist, allgemein zu reden, das Vermögen der Erkenntnisse. Diese bestehen in der bestimmten Beziehung gegebener Vorstellungen auf ein Objekt. Objekt aber ist das, in dessen Begriff das Mannigfaltige einer gegebenen Anschauung vereinigt ist. Nun erfordert aber alle Vereinigung der Vorstellungen Einheit des Bewusstseins in der Synthesis derselben." (B 137). Looking from our operational point of view we see that the mechanism described in this passage is that of "Vereinigung" (unification) of what we call "Vorstellungen" (mental constructions). Kant specifies further that this unification mechanism should act when "Vorstellungen" are synthesized ("in der Synthesis") and that the result of its operation is an (experiential) object. From this I conclude that this description could be clearly interpreted as an operational explanation (a proposition of a generative mechanism) of the totally new conception of "object" that Kant needs in order to articulate his Copernican Revolution.

Tightly related to that explanation, a few lines further Kant gives one of his most famous (and rare) examples: "Um aber irgend etwas im Raume zu erkennen, z.B. eine Linie, muss ich sie ziehen, und also eine bestimmte Verbindung des gegebenen Mannigfaltigen synthetisch zu Stande bringen, so, dass die Einheit dieser Handlung zugleich die Einheit des Bewusstseins (im Begriff einer Linie) ist, und dadurch allererst ein Objekt (ein bestimmter Raum) erkannt wird." [10] (B 138). Here Kant wants to explain a "line" and if we look at the functional side of what he writes we see that he presents an operational explanation of the kind defined earlier in this paper. In fact Kant describes here a "line mechanism" that, if you let it operate in your mind, gives you as result of your mental operations that "line" that he wants to explain. What is exemplified here for the case of a line is also paradigmatically valid for any experience-independent Element (a priori concepts like begin, end, singular, plural, unit, element, part, whole, space, time, number, here, there, now, etc.): with the help of Kant's Analytic of Concepts we should try to develop also for them generative mechanisms which, if we let them operate in our mind, would deliver the a priori concept that we want to explain.

What is now the relevance to robot design of such considerations? I think that they can help to answer the crucial question of experience (learning): What, if anything, should a robot contribute of its own to its knowledge ? In the light of Kant's explanations the answer is that what a robot should contribute of his own could be the mechanisms and operations of acquisition and application of experience-independent Elements ('reine Verstandesbegriffe a priori') that implement the unity of experience (categories) and are the conditions of possibility of experience.

This raises the question of how to investigate experience-independent Elements (a priori knowledge): what kind of elements are these, what role do they play in basic cognitive tasks, what functions are required for the performance of these tasks and what mechanisms can fulfill these requirements?  In the introduction to the Analytic Kant speaks of the main requirement to be fulfilled by his methods which is to "follow up the pure concepts to their first seeds and dispositions in the human understanding" (B 91) and presents then a short sketch of the method itself (B 90). His procedure will not be that of dissecting the content of concepts as usual in philosophical investigations; rather, he wants to apply a rarely attempted analytic procedure which consists in the dissection (modeling) of the faculty of the understanding itself.

After Kant, the operational method of dissecting the faculties and using the resulting architectural model as a foundation for metaphysical reflections has been abandoned in the mainstream philosophical discourse. One main reason was presented by Hegel [11] in his refutation of this Kantian method as "widersinnig" (absurd) at the beginning of his Introduction to the "Phänomenologie des Geistes": "Denn ist das Erkennen das Werkzeug, sich des absoluten Wesens zu bemächtigen, so fällt sogleich auf, dass die Anwendung eines Werkzeugs auf eine Sache sie vielmehr nicht lässt, wie sie für sich ist, sondern eine Formierung und Veränderung mit ihr vornimmt" [11, p.68]. Hegel's alternative is what could be called an "ontological method", which helps to explain "dasjenige, was an sich ist" and attain the goal of knowing the Absolute ("des Absoluten sich bemächtigen", [11, p.68]).

What should be noticed here in relation to robotics is that Kant’s original method of "dissecting the faculties" has later become a typical method of cybernetics, the science of operators, operations and operands. Similar operational methods have been successfully applied to a great variety of disciplines such as neuropsychology, machine design, computer science, etc. [1, Bonn 1991, p.724]. This provides another justification for reading in operational terms Kant's arguments for his fundamental hypothesis and for looking at this method as at a further pillar of a Kantian approach to robot engineering.

As Michael Oberhausen remarks [12]: "To simply consider Kant's apriori as innate still is a widespread cliché". This cliché is being disseminated not only by contemporary philosophers - with whom Kant is known to be in contradiction - but also by scientists when they approach Kant's work. Indeed, this is not astonishing if we consider that Kant's doctrine of the acquisition of experience-independent elements 'up to now has been quite ignored by research' [12, p. 6]. Fortunately we have now Oberhausen's comprehensive study which supplies the proof that Kant's doctrine of the acquisition of a priori constructs ('Vorstellungen') is not of no or only of secundary relevance but "lies at the center of critical philosophy". In his summary Oberhausen emphasises that Kant "in many different places of his work again and again explicitely explained that both space and time as well as pure concepts are acquired ..." [12, p.6]. The consequence for an autonomous robot are highly relevant: the robot must acquire not only its empirical knowledge but also the instruments, that alone can make experience possible!

In this sketch of a cybernetic approach to Kant's Analytic of Concepts, one adapted to the needs of robot engineering we make the following major claims and suggestions:

• that a priori concepts are necessary in autonoumous robots if they must be able to make experiences (to acquire a world model) in order to perform complex cognitive tasks;
• that Kant's theory of experience (in the Analytic of Concepts) could be used for developing a module that would allow a robot to make experiences thus acquiring its world model;
• that in order to apply it to robot engineering Kant's theory of experience must be read in an operational way;
• that it is possible and justified by Kant himself to read the Analytic of Concepts in an operational way
• that a robot should not only acquire its world model but also acquire the experience-independent elements - the condition of possibility of its experience.

What distinguishes our 'engineering reading' from classical 'philosophical readings' of Kant can be summarized as follows:

1. the principle of 'operational interpretation' allows to disentangle the functional from the metaphysical side of the Critique;
2. the use of that principle is not only justified based on Kant's distinction of 2 sides of the Critique but also necessary, if one wants to further develop and use Kant's functional model of mind in the context of robot engineering;
3. Kant's conception of the relation between "Vorstellung" and "Gegenstand" (Copernican Revolution) must be put at the center of our engineering reading because an autonomous robot is an 'artificial observer';
4. in Kant's Copernican Revolution we must focus on the functional instead of the metaphysical side: i.e., we must develop from scratch an operational interpretation of each single component of the hypothesis that "[objects] [must conform to] [our knowing]".
5. We consider Kant’s method of dissecting the faculties of mind as a typical cybernetic method that, although largely abandoned in philosophy, should be preferred to Hegel’s "ontological method" within the scope of robot engineering.

Basel, April 4, 2000.

  ©2000, Marco C. Bettoni, FHBB - 4.4..00 - 22.12.00