Aristotelian Intelligence
Perception is all you need.
I just finished reading Deborah Modrak’s excellent (and difficult) book Aristotle: The Power of Perception. Modrak’s project in that book is to demonstrate that one can extract a coherent philosophy of mind and a coherent epistemology from the Aristotelian corpus, and that both are based on Aristotle’s integrated model of perceptual activity and (in the case of humans) thinking, or noetic, activity.
Apart from the book’s contribution to our understanding of Hellenistic philosophy of mind (Julia Annas’ book on the philosophy of mind of the Stoics and the Epicureans should be mentioned here as well), it is also valuable as a conceptual framework for thinking about artificial intelligence (or, more narrowly, about the interplay between perception, pattern recognition, learning, and knowledge). Viewed from this angle, Aristotle’s thought looks remarkably modern and can offer some new insights for both analysis and synthesis of AI systems.
Generative and architectural thinkers
It will be useful to frame the discussion that follows in terms of the dichotomy between generative vs. architectural style of reasoning, proposed in the context of AI by Philip Agre in Computation and Human Experience. According to Agre,
generative theories … involve mathematical operations capable of generating an infinite number of mental structures by the repeated application of a small number of basic rules
while
architectural reasoning attempts not merely to implement a given abstraction but to discover abstractions that both do the required work and admit of natural implementations.
Herbert Simon, Alan Newell, and Noam Chomsky are generative thinkers, whereas Marvin Minsky (when he wrote The Society of Mind) is an architectural thinker. The core distinction between generativists and architecturalists hinges on the role of abstraction:
The generative and architectural styles of research have led to different views about the relationship between abstraction and implementation in human cognition. Generative theories posit large, consistent abstractions that operate by their own formally specifiable laws. McCarthy and the other proponents of the reconstruction of human knowledge in formal logic, for example, take as their starting point the generative power of logical formalisms and the formal consistency presupposed by their semantics.
…
Architectural theories, by contrast, hold that the demands of physical implementation have profound consequences for the formulation of abstract theories of cognition. Minsky in particular emphasizes physical locality so strongly that cognition becomes a large, fragmentary collection of mutually inconsistent abstractions, each bearing a different relationship to its physical implementation. He formulates his theory not as a single unified mechanism, or even as a single list of axioms of design, but as a constellation of mini-theories, each examining some feature of human intelligence - whether language, reasoning, decision-making, memory, emotion, imagination, or learning - on the level of engineering intuition.
The same tension can be found in the twentieth-century philosophy of science, where logical empiricists (a.k.a. logical positivists) are firmly in the generativist camp, whereas the proponents of the so-called New Mechanist Philosophy are architecturalists. Viewed through historical lens, this distinction makes perfect sense: The positivists of the Vienna circle, whose influence on later logical empiricists like Hempel was profound, took physics as their model science; by contrast, the New Mechanists (Bechtel, Craver, Darden) drew their inspiration primarily from the life sciences, specifically molecular biology and neuroscience. Newtonian physics, as interpreted by logical empiricists, is a paradigmatic generative framework: Everything in the Universe can be explained through consistent application of the laws of physics to the given initial condition. By contrast, biology and neuroscience are more concerned with how living organisms and living brains produce their activity. William Wimsatt is another example of architectural thinking in philosophy of science. I will return to the New Mechanists later, after introducing Modrak’s synthesis of Aristotle’s perception-based framework. For now, let’s tentatively agree to place Aristotle in the architectural camp.
Aristotle’s foundational principles
Modrak begins by stating, in modern terms, what she believes are five foundational principles underlying Aristotle’s theory of perception. These principles, comprising three descriptive principles and two prescriptive ones, “inform Aristotle’s treatment of the perceptual faculty and account for its unity and coherence.” Here they are, in Modrak’s words:
A. Descriptive Principles
Psychophysical Principle. Many states, if not all, that are ordinarily assigned to the soul are psychophysical states, namely, psychical states with physical realizations.
Actuality Principle. A cognitive faculty is potentially what its object is actually.
Sensory Representation Principle. If a cognitive activity has a sense object as its focal object, the psychic faculty involved is a perceptual faculty.
B. Prescriptive Principles
Analytic Principle. A psychological explanation should begin with an account of the constituent parts of the phenomenon under consideration and then make this account the basis for extending the explanation to cover more complex phenomena of the same sort.
Normative Psychophysical Principle. Psychological explanation at its most complete will take the psychophysical character of psychological states into account.
The key notion associated with Aristotle is hylomorphism, or the organizational unity of matter (hyle) and form (morphe). The above foundational principles instantiate this doctrine in the context of perception, referring to Aristotle’s idea that the soul (psyche) gives form to (or enforms) the body, as well as to the interplay between formal and material causes. Aristotle’s project is ambitious: Starting with the five senses, he wants to develop a cohesive account of perception, including apperception, or the ability to synthesize complex judgments based on multiple senses; imagination or phantasia, an umbrella term under which he includes sensory illusions, memory, dreaming, voluntary action, and even some forms of discursive reasoning; and, finally, the rational faculty, or the noetikon, which he attributes exclusively to humans. This emphasis on not just what the perceptual activity is, but also on how it is produced and what role different senses and faculties play in its production, is what makes Aristotle an architectural thinker, rather than a generativist like his mentor Plato.
On Modrak’s account, the Psychophysical Principle tells us that psychophysical states have both a functional description as mental states and a physical realization as states of the central nervous system. However, Aristotle’s functional description is closer to that of the biologists’ rather than that of the cognitive scientists’ since his psychophysical account does not decouple the functional view from the physical realization view. This is consistent with the hylomorphism doctrine and is very different from the multiple realizability ideas in functionalist accounts of cognition. The Actuality Principle and the Sensory Representation Principle together constitute a form of realism about sensory representations, but Aristotle views these representations as icons in the semiotic sense, i.e., as signs that stand for objects through relation of similarity.
The two prescriptive principles are explicitly architectural in nature. The Analytic Principle asks for explanations in terms of mechanisms that are constructed from basic constituent parts to form more complicated arrangements (e.g, when Aristotle begins with the analysis of the five basic senses and then proceeds to build up an explanatory framework for more complex perceptual phenomena). However, the psychophysical nature of the mental states has to be emphasized throughout, as demanded by the Normative Psychophysical Principle, so the resulting account does not amount to reducing mental states to nothing but neurophysiology (as eliminativist accounts along the lines of Paul and Patricia Churchland would do).
Multimodal perception and koine aisthesis
Aristotle’s account of the senses centers on the idea of logos, or rational mean. Sensory representations are given relative to opposing pairs (e.g., quiet vs. loud, dark vs. bright, bitter vs. sweet). The processing of composite percepts, when multiple sensory modalities provide information about a given object, is the function of what Aristotle called koine aisthesis, or the common sense. In accordance with the Analytic Principle, Aristotle aims for economy of explanation and avoids introducing another faculty where a perceptual faculty would do the work. Moreover, in line with his empiricist epistemology (e.g., he argues that knowledge is impossible without perception), learning plays a role in forming apperceptual judgments or in using the input of one sense to infer the input of another. For example, a stable association of “white” and “sweet” could arise in certain contexts — if I see a white object presented on a dessert plate and if I had seen and eaten similar-looking desserts in the past, then I may reliably infer that this object will be sweet to the taste without having tasted it.
In some way, this view of multiple sensory modalities acting in concert and reinforcing each other, with the aid of experience and learning, is reminiscent of Gerald Edelman’s ideas on neuronal group selection — assemblies of neurons that strongly respond to certain sensory modalities will be differentially reinforced as the organism acquires more experience; moreover, simultaneous presentation of multimodal sensory data about a given object will lead to reinforced connections between the neural assemblies corresponding to these different modalities (what Edelman refers to as re-entrant connections). Similar discussion of coherence among the senses and the unity of perception can also be found in the writings of Alain Berthoz, who argues for the importance of anticipation and prediction in addition to integrating the current sensory data into a whole, as well as for the (still not well-understood) role of simultaneous activity in many different regions of the brain and of temporal synchronization. Emery Brown’s description of general anesthesia in terms of synchronization and desynchronization of neural oscillations among different subsystems in the brain seems relevant here as well.
Imagination, or phantasia
Aristotle wants to explain how perceptual activity is produced and how it contributes to the functioning of the organism in its environment. As he writes in De Anima,
the soul of animals has been defined by two faculties, the faculty of discrimination (kritikon), which is the function of thought and perception, and the faculty that originates movement with respect to place.
Being an empiricist, Aristotle acknowledges that not every instance when these faculties are exercised involves perception of an external object — for example, when we anticipate or plan or dream. Guided by the Analytic Principle, Aristotle still associates perceptual objects with these types of activities, the phantasmata that are generated by the imaginative faculty or phantasia. As I pointed out earlier, Aristotle makes phantasia do a lot of work. To assimilate imagination into Aristotle’s general framework, Modrak formulates the following important idea:
Aristotle includes in the full formal description of a sensory experience the conditions under which the experience occurs. Perception occurs under standard conditions. Phantasia occurs under nonstandard conditions, that is, under conditions that are not conducive to veridical perception.
This covers sensory illusions, which arise when coherence among multiple senses breaks down; dreams; mental simulation and planning before taking an action (e.g., deciding where and how to move to avoid a predator); and certain forms of discursive reasoning that are nonpropositional in nature. All of these, according to Aristotle, involve manipulation of phantasmata, and it is not hard to see echoes of this in modern neurophysiological theories (e.g.., the role of efference copy in planning and coordinating movement) or in Daniel Dennett’s discussion of Skinnerian, Popperian, and Gregorian creatures which differ in their capabilities to perform internal simulations, to form hypotheses and conjectures, and to recruit tools in their external environment for the purpose of planning and acting. Memory is also an instance of phantasia, where we recall that a particular sensory impression occurred in the past.1 In line with the architectural style of reasoning, we can conceptualize both perception under standard conditions and the exercise of phantasia as production, manipulation, and exchange of icons (signs standing for objects through relation of similarity) and indices (signs standing for objects by being their causal effects) among different modules and architectural layers in the brain or in the AI system, as done, for example, by Manuel DeLanda.
Thought, or the noetic faculty
In Aristotelian epistemology, both the external objects of perception and their psychophysical correlates are particulars. Each of them is tied to a concrete instantiation either as part of objective external reality (Aristotle was a realist about objects of perception under standard conditions) or as a concrete psychophysical state. The internal sensory representations Aristotle has in mind are nonsymbolic, and both humans and nonhuman animals have the ability to manipulate them as they interact with their environment. However, humans, according to Aristotle, have the unique ability for rational thought, which involves manipulation of universals by the noetic faculty. Modrak explicitly says that “the mode of representation employed by the noetic faculty is properly described as symbolic.” However, these symbolic representations cannot exist by themselves, but require nonsymbolic (specifically, phantasmatic) vehicles. For example, when a geometer contemplates a problem involving triangles, the underlying thought process makes use of images of triangles as well as a considerable degree of abstraction:
The geometer ignores particularizing features such as size in order to treat the drawing as an arbitrarily selected instance of its class. Similarly, an image when used to facilitate reasoning about a type is treated as a representation of a token of the type under consideration, and the thinker ignores the features that are peculiar to it. The symbolic employment of a phantasma is already a distinct mode of representation. Moreover, there is some reason to believe that Aristotle envisages a kind of internal language as the vehicle for noetic representation. According to the De Interpretatione, ‘Spoken words are symbols of affections in the soul and written words are symbols of spoken words’’. The attribution of a quasi-linguistic character to noetic representation would also explain the persistent association of noetic capacities with ratiocination.
Aristotle’s noetikon is a manipulator of sentences in “mentalese,” which crucially relies on the perceptual faculty; this is exactly what Dennett had in mind when he spoke of theorem provers wrapped in an overcoat of transducers and effectors in The Intentional Stance. The universals figure prominently in Plato’s thought as well; however, what distinguishes Aristotle from Plato is that, for him, the universals (or syntactic objects, as we would say now) are formed in two ways: By induction from the particulars, which requires perception, or by valid logical inference, which requires ratiocination. Aristotle uses a curious military metaphor to describe the formation of universals out of a corpus of particular perceptions in The Posterior Analytics:
[O]ut of perception memory, as we say, comes to be and out of memory occurring frequently with respect to the same thing experience comes to be. For memories which are many by number make up one experience. And out of experience or out of the universal resting as a whole in the soul, the one beside the many, which is one and the same in all of these, comes to be the first principle of art and science— of art, if it concerns coming to be, of knowledge if it concerns being.
Neither are these present in us as definite dispositions nor do they come about from other dispositions which are more capable of knowing but from perception. Just as in a battle, when a rout occurs, if one takes a stand, and another takes a stand, then another, until the original formation is achieved. And the soul is such as to be capable of being affected in this way.
The first paragraph is almost Humean in spirit, arguing for the pragmatic inference of necessary connection from constant conjunction. The second paragraph, with the invocation of a rout in a battle, almost gestures at the Hebbian mechanism for learning. The overall picture that emerges is reminiscent of Aaron Sloman’s Virtual Machine Functionalism, where the notion of “virtual” refers not to virtual reality, but to virtualization, or the formation of manipulable objects and mechanisms on higher layers of a complex architectures while respecting the constraints imposed by lower layers (constraints that deconstrain). In particular, we can view the formation of symbolic representations (or universals) as virtualization built on iconic-indexical substrates, which in turn can be instantiated as an integrated activity of a number of perceptual mechanisms. Yet another piece of evidence in favor of viewing Aristotle as an architectural thinker!
New Mechanist Philosophy and the importance of architecture
I will close by mentioning a close link between Modrak’s modern take on Aristotelian hylomorphism and the New Mechanist Philosophy. A good source for this is Daniel De Haan’s “Hylomorphism and the New Mechanist Philosophy in Biology, Neuroscience, and Psychology.” The proponents of New Mechanism are interested in showing how a given phenomenon is produced by its causes. New Mechanistic explanations are not reductive in the usual sense of “X is nothing but Y;” instead, they posit the mosaic view of science (or what Sunny Auyang in her book on complex system theories called “federal unity of science” and contrasted it with the “imperial view” of the reductionists). As Carl Craver writes in his 2007 book Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience,
the mosaic view treats the unity of science as the collaborative accumulation of constraints at multiple levels. Whereas reduction focuses on relations of identity, supervenience, and ontological reductive links, the mechanistic mosaic view emphasizes the importance of explanatory relevance as the bridge between levels. Finally, whereas reduction models emphasize the importance of explanatory reduction to fundamental levels, the mosaic view can be pluralistic about levels, recognizing the genuine importance of higher-level causes and explanations. The mosaic unity of science is constructed during the process of collaboration by different fields in the search for multilevel mechanisms.
This is the essence of the architectural style of reasoning about complex systems, although here I agree with John Doyle about the importance of not confusing levels with layers. Levels pertain to a hierarchy arranged by scale, whereas layers describe functional organization. William Wimsatt uses the term “perspectives” to refer to layers in this sense. A given layer can be described on multiple levels (e.g., we could look at a particular region in the brain and look at its organization as an assembly of neurons, then look at the scale of individual neurons; etc.). Architectural explanations move across levels and layers. In this sense, Aristotle’s integrated view of perception, action, thought, and knowledge is a superb example of an architectural theory, and may yet show its usefulness in the context of AI as some researchers are starting to contemplate moving away from training on language and towards the new era of experience.
This touches upon Aristotle’s distinction between mneme, or memory, and anamnesis, or recollection. The former is just a form of pattern recognition, while the latter is a conscious act, a deliberate search for an object perceived or learned in the past. Modrak argues that, while Aristotle is vague on how anamnesis can be fit into a purely perceptual framework, it is plausible to view it as an emergent property of the human perceptual capacity, arising due to its organizational complexity. Gerald Edelman’s distinction between primary and secondary consciousness could be interpreted in the same way.