Brendan Fleig-Goldstein

This paper presents a topological learning-theoretic analysis of an approach in cognitive science called bounded rational analysis. In this approach, modelers begin by deriving an optimal cognitive model, then use discrepancies between idealized calculations and observed human behavior to identify psychological constraints (e.g., memory limits), and incorporate such constraints into a newly derived resource-optimal model in an iterative scientific process. I show that this de-idealization process exploits an epistemic asymmetry: models positing greater rationality are more falsifiable. A methodological preference for more rational models can therefore be epistemically justified via learning theory.

There is a general view that scientific inferences from data to theory are rarely, if ever, deductive inferences. Against this view is a rising tide of work showing the importance of deductive inference in science. Such inferences have been called demonstrative inductions (DI). However, critics have argued against the epistemic advantages of DI, claiming that it merely passes the buck of inductive risk to the premises. In this paper, I defend the advantages of demonstrative induction. I argue that the virtues of structuring inferences deductively derive mainly from the ability to ``isolate'' inductive risk to specific areas of theory, as well as from the ability to ``export'' this risk to more certain domains than the one under investigation. I support this argument with examples from cognitive science, where I argue that the current context of inquiry makes non-deductive, hypothesis testing particularly ineffective. Properly characterized bounded rationality assumptions can, however, help facilitate DI, allowing cognitive scientists to derive process models from environmental structure, export inductive risk outside the head, and isolate the remaining inductive risk to claims about cognitive limitations. I show how this strategy ameliorates the underdetermination problems that plague hypothesis testing, especially in early investigations of complicated black-box systems.

Why might a scientist want to establish a cognitive model as optimally suited to some particular environment? In this paper, I suggest that an unexamined motivation for establishing models as optimal is to uncover systematic discrepancies between idealized human behavior and observed human behavior. These discrepancies can lead to the discovery of previously unknown cognitive architecture details (e.g., resource constraints), which can then be incorporated into models and give rise to new idealized models that factor in these newly uncovered details. Further discrepancies then arise, and the process repeats itself in an iterative fashion. Most importantly, each step in this process provides evidence for descriptive claims about human cognition. The point, then, of establishing optimal models is to facilitate a particular process for marshalling evidence about the human cognitive system.

What capacities must an AI system possess to be held responsible for its actions? I argue that AI systems can be accountable agents when they possess sufficiently strong commitments to relevant norms (ethical, rational, or conventional). This paper articulates empirically determinable necessary and sufficient conditions for possessing such commitments. Specifically, I argue that what I term a meta-reflective capacity toward a goal is both necessary and sufficient. Meta-reflection is the capacity to maintain resource-optimal performance by appropriately changing one’s cognitive strategy in response to changes in internal constraints. A commitment’s strength can be characterized by the constraints under which a system fails to maintain resource-optimality. The path forward for a theory of AI responsibility requires articulating which qualitative internal system limitations are excusable, forgivable, or competence-undermining. Considerations of the mutability of constraints offer a partial way to delineate such classes. This framework connects philosophical theories of responsibility to cognitive processes and provides a path toward identifying and engineering normative commitments in biological and artificial systems.

This paper argues that Newton employed a non-hypothetical method of evidential reasoning called demonstrative induction in his chronological studies. In demonstrative induction, inductive risk is confined to the premises, and the secureness of the conclusion depends on the secureness of those premises. I show that Newton's approach to chronology exemplifies two key restrictions on demonstrative induction: (i) premises must themselves be supported by demonstrative induction or by inductive generalizations, and (ii) inductive risk should be confined to premises drawn from domains of inquiry whose claims are better established, thereby “exporting” inductive risk to more certain domains. Drawing on Newton's use of sources from history, biology, social science, and astronomy, I analyze how he transformed data from these fields into evidence for his chronological scheme. The strategy of exporting inductive risk, I argue, reveals a common evidential pattern across his natural-philosophical, historical, and theological work, and sheds light on the deeper methodological unity that underlies his otherwise disparate intellectual pursuits.