Connectionism and Compositionality

Recent debates over artificial intelligence have focused on its perceived lack of interpretability and explainability. I argue that these notions fail to capture an important aspect of what end-users—as opposed to developers—need from these models: what is needed is systematicity, in a more demanding sense than the compositionality-related sense that has dominated discussions of systematicity in the philosophy of language and cognitive science over the last thirty years. To recover this more demanding notion of systematicity, I distinguish between (i) the (...) systematicity of thinkable contents, (ii) the systematicity of thinking, and (iii) the ideal of systematic thought. I then deploy this distinction to critically evaluate Fodor’s systematicity-based argument for the language of thought hypothesis before recovering the notion of the systematicity of thought as a regulative ideal, which has historically shaped our understanding of what it means for thought to be rational, authoritative, and scientific. To assess how much systematicity we need from AI models, I then argue, we must look to the functions of systematizing thought. To this end, I identify five functions served by systematization, and show how these can be used to arrive at a dynamic understanding of the need to systematize thought that can tell us what kind of systematicity is called for and when. (shrink)

In his famous 1982 paper, Allen Newell [22, 23] introduced the notion of knowledge level to indicate a level of analysis, and prediction, of the rational behavior of a cognitive arti cial agent. This analysis concerns the investigation about the availability of the agent knowledge, in order to pursue its own goals, and is based on the so-called Rationality Principle (an assumption according to which "an agent will use the knowledge it has of its environment to achieve its goals" [22, (...) p. 17]. By using the Newell's own words: "To treat a system at the knowledge level is to treat it as having some knowledge, some goals, and believing it will do whatever is within its power to attain its goals, in so far as its knowledge indicates" [22, p. 13]. In the last decades, the importance of the knowledge level has been historically and system- atically downsized by the research area in cognitive architectures (CAs), whose interests have been mainly focused on the analysis and the development of mechanisms and the processes governing human and (arti cial) cognition. The knowledge level in CAs, however, represents a crucial level of analysis for the development of such arti cial general systems and therefore deserves greater research attention [17]. In the following, we will discuss areas of broad agree- ment and outline the main problematic aspects that should be faced within a Common Model of Cognition [12]. Such aspects, departing from an analysis at the knowledge level, also clearly impact both lower (e.g. representational) and higher (e.g. social) levels. (shrink)

The systematicity argument only challenges connectionism if systematicity is a general property of cognition. I examine this thesis in terms of properties of concepts. First, I propose that Evans's Generality Constraint only applies to attributions of belief. Then I defend a variety of conceptual pluralism, arguing that concepts share two fundamental properties related to centrality and belief-attribution, and contending that there are two kinds of concepts that differ in their compositional properties. Finally, I rely on Dual Systems Theory and on (...) differences between animal and human cognition to suggest a scenario of two processing systems that work on different kinds of concepts, with only one of them supporting full systematicity. I sketch a non-classical systematicity argument that rules out classicism as the basis of one of those systems given that it would wrongly entail that both systems are fully systematic. (shrink)

A comprehensive introduction to the Language of Though Hypothesis (LOTH) accessible to general audiences. LOTH is an empirical thesis about thought and thinking. For their explication, it postulates a physically realized system of representations that have a combinatorial syntax (and semantics) such that operations on representations are causally sensitive only to the syntactic properties of representations. According to LOTH, thought is, roughly, the tokening of a representation that has a syntactic (constituent) structure with an appropriate semantics. Thinking thus consists in (...) syntactic operations defined over representations. Most of the arguments for LOTH derive their strength from their ability to explain certain empirical phenomena like productivity, systematicity of thought and thinking. (shrink)

Fodor and Pylyshyn's critique of connectionism has posed a challenge to connectionists: Adequately explain such nomological regularities as systematicity and productivity without postulating a "language of thought" (LOT). Some connectionists like Smolensky took the challenge very seriously, and attempted to meet it by developing models that were supposed to be non-classical. At the core of these attempts lies the claim that connectionist models can provide a representational system with a combinatorial syntax and processes sensitive to syntactic structure. They are not (...) implementation models because, it is claimed, the way they obtain syntax and structure sensitivity is not "concatenative," hence "radically different" from the way classicists handle them. In this paper, I offer an analysis of what it is to physically satisfy/realize a formal system. In this context, I examine the minimal truth-conditions of LOT Hypothesis. From my analysis it will follow that concatenative realization of formal systems is irrelevant to LOTH since the very notion of LOT is indifferent to such an implementation level issue as concatenation. I will conclude that to the extent to which they can explain the law-like cognitive regularities, a certain class of connectionist models proposed as radical alternatives to the classical LOT paradigm will in fact turn out to be LOT models, even though new and potentially very exciting ones. (shrink)

This paper offers both a theoretical and an experimental perspective on the relationship between connectionist and Classical (symbol-processing) models. Firstly, a serious flaw in Fodor and Pylyshyn’s argument against connectionism is pointed out: if, in fact, a part of their argument is valid, then it establishes a conclusion quite different from that which they intend, a conclusion which is demonstrably false. The source of this flaw is traced to an underestimation of the differences between localist and distributed representation. It has (...) been claimed that distributed representations cannot support systematic operations, or that if they can, then they will be mere implementations of traditional ideas. This paper presents experimental evidence against this conclusion: distributed representations can be used to support direct structure-sensitive operations, in a man- ner quite unlike the Classical approach. Finally, it is argued that even if Fodor and Pylyshyn’s argument that connectionist models of compositionality must be mere implementations were correct, then this would still not be a serious argument against connectionism as a theory of mind. (shrink)

We describe and try to motivate our project to build systems using both a knowledge based and a neural network approach. These two approaches are used at different stages in the solution of a problem, instead of using knowledge bases exclusively on some problems, and neural nets exclusively on others. The knowledge base (KB) is defined first in a declarative, symbolic language that is easy to use. It is then compiled into an efficient neural network (NN) representation, run, and the (...) results from run time and (eventually) from learning are decompiled to a symbolic description of the knowledge contained in the network. After inspecting this recovered knowledge, a designer would be able to modify the KB and go through the whole cycle of compiling, running, and decompiling again. The central question with which this project is concerned is, therefore, How do we go from a KB to an NN, and back again? We are investigating this question by building tools consisting of a repertoire of language/translation/network types, and trying them on problems in a variety of domains. (shrink)