(Accepted at Mind and Language) Intertemporal Rationality Without Temporal Representation
ABSTRACT: Recent influential accounts of temporal representation—the use of mental representations with explicit temporal contents, such as before and after relations, durations, and specific times—have sharply distinguished representation from mere sensitivity. A common, important picture of intertemporal rationality is that it consists in appropriately trading off immediate and future rewards, to maximize total expected discounted utility across time. By carefully analysing simple reinforcement learning algorithms, this paper shows that, given such notions of temporal representation and intertemporal rationality, it would be possible for an agent to achieve intertemporal rationality without temporal representation. Therefore, either: the austere account of temporal representation is too demanding; the utility-maximizing account of intertemporal rationality is not demanding enough; or the relationship between intertemporal rationality and temporal representation is very different to what many have assumed.
(Accepted at Behavioral and Brain Sciences) Positing numerosities may be metaphysically extravagant; positing representation of numerosities is not.
(Commentary on S. Clarke & J. Beck 'The Number Sense Represents (Rational) Numbers')
ABSTRACT: Clarke & Beck assume that ANS representations should be assigned referents from our scientific ontology. However, many representations, both in perception and cognition, do not straightforwardly refer to such entities. If we reject Clarke and Beck’s assumption, many possible contents for ANS representations besides number are compatible with the evidence Clarke & Beck cite.
Narratives, Reinforcement Learning, and Death in Other Animals
ABSTRACT: I distinguish two broad pictures of rationality when it come to decisions which will have effects into one's future. On one picture, the rational such choice in such cases is the one which maximizes expected utility over time, and our ordinary decision-making imperfectly approximates such rationality even though we do not (and in many cases cannot) think in terms of the relevant mathematical formalism. On the rival picture, rational choice corresponds more to how ordinary humans make decisions, and can take into account factors which go beyond utility maximization — specifically, how actions contribute to the overall narrative structure of one's life. On this latter view, there is a sharp difference between humans and those animals who are incapable of thinking in terms of narratives or extended temporal structures. I defend the former view of rationality by showing how it can explain away intuitions about the importance of narratives which appear to favour the latter view. Reinforcement learning and narratives can be thought of as complementary tools — each imperfect — for approximating utility maximization. Specifically, RL algorithms converge on utility maximization under many conditions, but in environments with non-Markovian dynamics, they may break down. Narratives, I argue, are perfectly suited to helping a system imperfectly approximate utility maximization in such non-Markovian environments. This view does justice to important differences between humans and animals when it comes to decision-making, whilst avoiding positing a fundamental divide. It thereby undermines arguments (e.g. by Velleman) which try to derive the claim that painless death is not a harm to other animals from the alleged existence of such a divide.
Temporal Competence and Temporal Representation
ABSTRACT: What is temporal representation—the use of mental representations with explicit temporal contents, such as before and after relations, durations, and specific times? Many theories of representation and discussions of animals' behaviour with respect to time seem to imply that temporal representation is extremely widespread, as nearly all organisms are in some way temporally competent — that is, they produce behaviour at the right times given the dynamics of their environment. However such temporal competence can usually be explained with very simple mechanisms, and describing them as 'representing' the temporal features in question does not add anything to simply describing those mechanisms. This has motivated some philosophers to give accounts of temporal representation which demand more of temporal representation than simple temporal competence. However, I show that extant versions of such accounts are too demanding, and appear to be motivated by the demand for behaviour which can only be explained by positing temporal representation — a demand which, I argue, is impossible to satisfy given that there will always be explanations which describe the implementation of representational processes. I give an alternative account of temporal representation which shows how temporal representations can be explanatory while also avoids excessive demandingness. This alternative develops the thought that temporal representation allows for flexibility with respect to time. To do this, I give a novel account of the relevant kind of flexibility, and then show how states which are coupled to a temporal feature X and which are apt to combine with other representational states ground that kind of flexibility.
Please Email for latest version.
Episodic Memory and Unrestricted Learning
ABSTRACT: We use rich memories for particular events every day, for all kinds of tasks. Need to remember a recipe, a fun fact about a certain species, or what a city you've visited many times before is like? You might well proceed by recalling a particular occasion when you made the dish in question, learned about the species, or visited the city. Why? Isn't recalling all this irrelevant detail a waste of resources, and sometimes actively harmful, given that we also have forms of memory which incorporate information from multiple different episodes, have less distracting details, and do not require so much sophistication to store, access, and operate with? I argue that this question poses a deep problem for a number of accounts of the function of episodic memory. I then draw on recent ideas in AI to show that we can answer this problem, and show a crucial role for episodic memory in intelligence generally, if we appreciate the role episodic memory can play in a special kind of learning. Whereas most kinds of learning have in principle limits to how much they can learn about the world, learning with access to episodic memory is unrestricted in this sense. If episodic memory is for unrestricted learning, we can give explanations of its use in a wide range of contexts which seem to have little in common, and to which other forms of memory seem at first glance to be better suited.
Draft. Old version (The Significance of Episodic Memory) Please Email for Latest Version
Cognitive Significance and Cross-Species Attribution of Episodic Memory
ABSTRACT: I argue for a new approach to defining episodic memory for cross-species comparison. When we find creatures with states sharing some features with typical human episodic memories—e.g. being about particular events, being realized by a neural area homologous to the hippocampus, and involving mental imagery—whilst lacking others—e.g. embedding in a narrative, or higher-order awareness that this is an event from my past—do such creatures have episodic memory or not? How should we even go about answering such questions—how do we determine which features matter? The scientific and philosophical literature tends to opt for extreme views, by requiring such states to have all of the features of episodic memory (in which case no other animals will count), or to simply have any features of human episodic memory (in which case episodic memory will be extremely widespread, but will appear in many different forms in different species). I argue for an alternative approach which systematically picks out a notion of episodic memory that is useful in a wide range of contexts, avoids the pitfalls of these other approaches, and can be extended to the individuation of many other mental states where similar issues arise for cross-species comparison. My approach revolves around seeking sets of features which interact with one another to jointly make a big difference to the overall capacities of the mind. Using imagery to allow for efficient storage of rich representations of particular past events allows for unrestricted learning, and is therefore one such cluster of jointly cognitively significant features. This view can systematize existing empirical research on episodic memory in different species, and can motivate new experimental paradigms for such research.
Old version (The Significance of Episodic Memory) Please Email for Latest Version
What If Anything is Semantic Memory Like in Non-Human Animals?
ABSTRACT: Much of the literature on animal memory focuses either on conditioning and reinforcement learning, or on whether animals have episodic memory. This latter literature tends to assume that animals do have semantic memory, treating episodic memory as a more advanced, mysterious and interesting capacity. I argue that the assumption that other animals have semantic memory only makes sense on an understanding of semantic memory as merely involving memory for generics as opposed to particulars (with contents more like Lions are dangerous than Gary the gazelle was eaten by a lion). But the notion of semantic memory appealed to in human psychology mixes this conception of semantic memory with at least one other conception: memory in a language-like format. On this alternative conception, it is much less clear which, if any, non-human animals have semantic memory. Furthermore, these points suggest that semantic memory does not form a natural kind at all: rather, it seems to lump together quite different forms of memory solely in virtue of their not being episodic. I explore new theoretical and empirical questions which are opened up by asking about the different forms of semantic memory, which animals have them, and why they matter.
Slides from presentation at Issues in Philosophy of Memory 2.5 Online/Grenoble, July 2021. Please Email for Latest Version