Recent progress in artificial intelligence has renewed interest in building systems that learn and think like people. Many advances have come from using deep neural networks trained end-to-end in tasks such as object recognition, video games, and board games, achieving performance that equals or even beats that of humans in some respects. Despite their biological inspiration and performance achievements, these systems differ from human intelligence in crucial ways. We review progress in cognitive science suggesting that truly human-like learning and thinking (...) machines will have to reach beyond current engineering trends in both what they learn and how they learn it. Specifically, we argue that these machines should build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems; ground learning in intuitive theories of physics and psychology to support and enrich the knowledge that is learned; and harness compositionality and learning-to-learn to rapidly acquire and generalize knowledge to new tasks and situations. We suggest concrete challenges and promising routes toward these goals that can combine the strengths of recent neural network advances with more structured cognitive models. (shrink)

'We desire all and only those things we conceive to be good; we avoid what we conceive to be bad.' This slogan was once the standard view of the relationship between desire or motivation and rational evaluation. Many critics have rejected this scholastic formula as either trivial or wrong. It appears to be trivial if we just define the good as 'what we want', and wrong if we consider apparent conflicts between what we seem to want and what we seem (...) to think is good. In Appearances of the Good, Sergio Tenenbaum argues that the old slogan is both significant and right, even in cases of apparent conflict between our desires and our evaluative judgements. Maintaining that the good is the formal end of practical inquiry in much the same way as truth is the formal end of theoretical inquiry, he provides a fully unified account of motivation and evaluation. (shrink)

Rational Powers in Action presents a conception of instrumental rationality as governing actions that are extended in time with indeterminate ends. Tenenbaum argues that previous philosophical theories in this area, in focusing on momentary snapshots of the mind of idealized agents, miss central aspects of human rationality.

Shepard has argued that a universal law should govern generalization across different domains of perception and cognition, as well as across organisms from different species or even different planets. Starting with some basic assumptions about natural kinds, he derived an exponential decay function as the form of the universal generalization gradient, which accords strikingly well with a wide range of empirical data. However, his original formulation applied only to the ideal case of generalization from a single encountered stimulus to a (...) single novel stimulus, and for stimuli that can be represented as points in a continuous metric psychological space. Here we recast Shepard's theory in a more general Bayesian framework and show how this naturally extends his approach to the more realistic situation of generalizing from multiple consequential stimuli with arbitrary representational structure. Our framework also subsumes a version of Tversky's set-theoretic model of similarity, which is conventionally thought of as the primary alternative to Shepard's continuous metric space model of similarity and generalization. This unification allows us not only to draw deep parallels between the set-theoretic and spatial approaches, but also to significantly advance the explanatory power of set-theoretic models. Key Words: additive clustering; Bayesian inference; categorization; concept learning; contrast model; features; generalization; psychological space; similarity. (shrink)

In many learning or inference tasks human behavior approximates that of a Bayesian ideal observer, suggesting that, at some level, cognition can be described as Bayesian inference. However, a number of findings have highlighted an intriguing mismatch between human behavior and standard assumptions about optimality: People often appear to make decisions based on just one or a few samples from the appropriate posterior probability distribution, rather than using the full distribution. Although sampling-based approximations are a common way to implement Bayesian (...) inference, the very limited numbers of samples often used by humans seem insufficient to approximate the required probability distributions very accurately. Here, we consider this discrepancy in the broader framework of statistical decision theory, and ask: If people are making decisions based on samples—but as samples are costly—how many samples should people use to optimize their total expected or worst-case reward over a large number of decisions? We find that under reasonable assumptions about the time costs of sampling, making many quick but locally suboptimal decisions based on very few samples may be the globally optimal strategy over long periods. These results help to reconcile a large body of work showing sampling-based or probability matching behavior with the hypothesis that human cognition can be understood in Bayesian terms, and they suggest promising future directions for studies of resource-constrained cognition. (shrink)

In this paper we advance a new solution to Quinn’s puzzle of the self-torturer. The solution falls directly out of an application of the principle of instrumental reasoning to what we call “vague projects”, i.e., projects whose completion does not occur at any particular or definite point or moment. The resulting treatment of the puzzle extends our understanding of instrumental rationality to projects and ends that cannot be accommodated by orthodox theories of rational choice.

Deontological theories face difficulties in accounting for situations involving risk; the most natural ways of extending deontological principles to such situations have unpalatable consequences. In extending ethical principles to decision under risk, theorists often assume the risk must be incorporated into the theory by means of a function from the product of probability assignments to certain values. Deontologists should reject this assumption; essentially different actions are available to the agent when she cannot know that a certain act is in her (...) power, so we cannot simply understand her choice situation as a “risk-weighted” version of choice under certainty. (shrink)

We present statistical analyses of the large‐scale structure of 3 types of semantic networks: word associations, WordNet, and Roget's Thesaurus. We show that they have a small‐world structure, characterized by sparse connectivity, short average path lengths between words, and strong local clustering. In addition, the distributions of the number of connections follow power laws that indicate a scale‐free pattern of connectivity, with most nodes having relatively few connections joined together through a small number of hubs with many connections. These regularities (...) have also been found in certain other complex natural networks, such as the World Wide Web, but they are not consistent with many conventional models of semantic organization, based on inheritance hierarchies, arbitrarily structured networks, or high‐dimensional vector spaces. We propose that these structures reflect the mechanisms by which semantic networks grow. We describe a simple model for semantic growth, in which each new word or concept is connected to an existing network by differentiating the connectivity pattern of an existing node. This model generates appropriate small‐world statistics and power‐law connectivity distributions, and it also suggests one possible mechanistic basis for the effects of learning history variables (age of acquisition, usage frequency) on behavioral performance in semantic processing tasks. (shrink)

It seems to be a humdrum fact of human agency that we act on intentions or decisions that we have made at an earlier time. At breakfast, you look at the Taco Hut menu online and decide that later today you’ll have one of their avocado burritos for lunch. You’re at your desk and you hear the church bells ring the noon hour. You get up, walk to Taco Hut, and order the burrito as planned. As mundane as this sort (...) of scenario might seem to be, philosophers have raised a problem in understanding it. If you are simply abiding by this morning’s decision, how are you acting autonomously? Your earlier self seems to be calling the shots; if you are just acting accordingly, without thinking through it or in some other way trying to ensure that the past decision conforms to your present standpoint, it is not clear how this amounts to an exercise of your present autonomous agency. It seems, rather, that your earlier self has succeeded in slaving you to her own purposes. She was the one who wanted (intended, judged it to be good, etc.) to have an avocado burrito. In simply following through, your current self seems to be just an automaton performing the commands left behind by your former self. Of course, you might not allow yourself to be shackled by your earlier self. You might refuse to follow anything but your own present judgments: you will only go to Taco Hut if this is what you judge you should do right now, and once at Taco Hut you will only eat the avocado burrito if that is what you want to eat once there. But if this is the way you generally operate, this seems to block your ability to make effective future-directed decisions. The puzzle, then, is one of explaining how the future self can do the bidding of her past self without losing her autonomy. We call this “the Problem of Diachronic Autonomy.” Philosophers raising this problem take it to show that there must be reasons or rational requirements to follow-through with our past decisions. According to these philosophers, we can only make sense of our diachronic autonomy if our past decisions put rational pressure on us to follow through. We argue that there is no Problem of Diachronic Autonomy. There is, in other words, no puzzling situation that needs explaining. Consequently, there is no need coming from this purported puzzle to think that our future-directed decisions generate reasons or rational requirements to follow through. The correct view of our diachronic autonomy is the “naïve” one: the “future self” can do the bidding of the “past self” without giving up its autonomy because, very simply, the past self is the same agent as the future self. I am acting autonomously when I get the avocado burrito, because I was the one who decided to get the burrito. I am acting on my own freely-formed decision. (shrink)

It is undeniable that human agents sometimes act badly, and it seems that they sometimes pursue bad things simply because they are bad. This latter phenomenon has often been taken to provide counterexamples to views according to which we always act under the guise of the good. This paper identifies several distinct arguments in favour of the possibility that one can act under the guise of the bad. GG seems to face more serious difficulties when trying to answer three different, (...) but related, arguments for the possibility of acting under the guise of the bad. The main strategies available to answer these objections end up either undermining the motivation for GG or failing to do full justice to the nature of perverse motivation. However, these difficulties turn out to be generated by focusing on a particular version of GG, what I call the “content version”. But we have independent reasons to prefer a different version of GG; namely, the “attitude version”. The attitude version allows for a much richer understanding of the possibility of acting on what we conceive to be bad. Drawing on an analogy with theoretical akrasia and theoretical perversion, I try to show how the attitude version can provide a compelling account of perverse actions. (shrink)

Hierarchical Bayesian models (HBMs) provide an account of Bayesian inference in a hierarchically structured hypothesis space. Scientific theories are plausibly regarded as organized into hierarchies in many cases, with higher levels sometimes called ‘paradigms’ and lower levels encoding more specific or concrete hypotheses. Therefore, HBMs provide a useful model for scientific theory change, showing how higher‐level theory change may be driven by the impact of evidence on lower levels. HBMs capture features described in the Kuhnian tradition, particularly the idea that (...) higher‐level theories guide learning at lower levels. In addition, they help resolve certain issues for Bayesians, such as scientific preference for simplicity and the problem of new theories. *Received July 2009; revised October 2009. †To contact the authors, please write to: Leah Henderson, Massachusetts Institute of Technology, 77 Massachusetts Avenue, 32D‐808, Cambridge, MA 02139; e‐mail: [email protected]. (shrink)

This paper argues that the principles of instrumental rationality apply primarily to extended action through time. Most philosophers assume that rational requirements and principles govern in the first instance momentary mental states, as opposed to governing extended intentional actions directly. In the case of instrumental rationality, the relevant mental states or attitudes would typically be preferences, decisions, or intentions. In fact, even those who recognize the extended nature of our agency still assume that rational requirements apply primarily to mental states (...) at a moment in time. Such views try to do justice to the extended nature of our agency by postulating rational requirements that apply in the first instance to plans, policies, and intentions more generally. The paper focuses on the central case of requirements and reasons governing the reconsideration of intentions and argues that these requirements or reasons are either superfluous or invalid.I argue that a proper conception of instrumental reasoning that applies directly to actions turn out to have surprising consequences. In fact, this conception allows us to see that policies, projects and the like are best understood as instances of extended actions, and that the instrumental requirements that apply to projects and policies are exactly the same as the instrumental requirements that apply to ordinary extended actions. Finally, I argue that the resulting theory of instrumental rationality is a significant improvement over theories that rely on principles governing intentions. (shrink)

Novice designers produced a sequence of sketches while inventing a logo for a novel brand of soft drink. The sketches were scored for the presence of specific objects, their local features and global composition. Self‐assessment scores for each sketch and art critics' scores for the end products were collected. It was investigated whether the design evolves in an essentially random fashion or according to an overall heuristic. The results indicated a macrostructure in the evolution of the design, characterized by two (...) stages. For the majority of participants, the first stage is marked by the introduction and modification of novel objects and their local and global aspects; the second stage is characterized by changes in their global composition. The minority that showed the better designs has a different strategy, in which most global changes were made in the beginning. Although participants did not consciously apply these strategies, their self‐assessment scores reflect the stages of the process. (shrink)

We present statistical analyses of the large‐scale structure of 3 types of semantic networks: word associations, WordNet, and Roget's Thesaurus. We show that they have a small‐world structure, characterized by sparse connectivity, short average path lengths between words, and strong local clustering. In addition, the distributions of the number of connections follow power laws that indicate a scale‐free pattern of connectivity, with most nodes having relatively few connections joined together through a small number of hubs with many connections. These regularities (...) have also been found in certain other complex natural networks, such as the World Wide Web, but they are not consistent with many conventional models of semantic organization, based on inheritance hierarchies, arbitrarily structured networks, or high‐dimensional vector spaces. We propose that these structures reflect the mechanisms by which semantic networks grow. We describe a simple model for semantic growth, in which each new word or concept is connected to an existing network by differentiating the connectivity pattern of an existing node. This model generates appropriate small‐world statistics and power‐law connectivity distributions, and it also suggests one possible mechanistic basis for the effects of learning history variables (age of acquisition, usage frequency) on behavioral performance in semantic processing tasks. (shrink)

In trying to explain the possibility of akrasia , it seems plausible to deny that there is a conceptual connection between motivation and evaluation ; akrasia occurs when the agent is motivated to do something that she does not judge to be good . However, it is hard to see how such accounts could respect our intuition that the akratic agent acts freely, or that there is a difference between akrasia and compulsion. It is also hard to see how such (...) accounts could be extended to the realm of theoretical reason, but this is generally not taken Ito be a problem, because it is generally assumed that there is no similar phenomenon in the realm of theoretical reason. This paper argues that there is such a thing as theoretical akrasia, and that we can find a characterization of this phenomenon in Descartes’s Meditations. Drawing on certain passages in the Meditations, we can construct an account of theoretical akrasia; this account can then be adapted to resolve the original problem of akrasia in the realm of practical reason. The account asserts that there is a conceptual connection between motivation and evaluation in free action; it also enables us to show how the akratic agent is still acting freely when he does something that he does not judge to be the best all things considered. (shrink)

Hierarchical Bayesian models (HBMs) provide an account of Bayesian inference in a hierarchically structured hypothesis space. Scientific theories are plausibly regarded as organized into hierarchies in many cases, with higher levels sometimes called ‘para- digms’ and lower levels encoding more specific or concrete hypotheses. Therefore, HBMs provide a useful model for scientific theory change, showing how higher-level theory change may be driven by the impact of evidence on lower levels. HBMs capture features described in the Kuhnian tradition, particularly the idea (...) that higher-level theories guide learning at lower levels. In addition, they help resolve certain issues for Bayesians, such as scientific preference for simplicity and the problem of new theories. (shrink)

Constitutivists have tried to answer Enoch’s “schmagency” objection by arguing that Enoch fails to appreciate the inescapability of agency. Although these arguments are effective against some versions of the objection, I argue that they leave constitutivism vulnerable to an important worry; namely, that constitutivism leaves us alienated from the moral norms that it claims we must follow. In the first part of the paper, I try to make this vague concern more precise: in a nutshell, it seems that constitutivism cannot (...) provide an adequate account of the relation between the constitutive norms of agency and the particular ends the agent pursues. I then provide a broad outline of an interpretation of Kant’s formalism that is immune to this objection. I conclude that constitutivism is best understood as the upshot of a formalist view of categorical practical principles. (shrink)

The idea of direction of fit has been found appealing by many philosophers. Anscombe’s famous examples have persuaded many of us that there must be some deep difference between belief and desire that is captured by the metaphor of direction of fit. Most of the aim of the paper is to try to get clear on which intuitions Anscombe’s example taps into. My view is that there is more than one intuition in play here, and I will try to show (...) that various distinctions and points are confused in the literature on direction of fit. But I also want to argue that once the proper distinctions are made, it’s not clear that the notion of direction of fit can do any of the philosophically significant work that it was supposed to do. I first argue that the best way to unpack the notion of direction of fit would indeed be by means of the constitutive relation between truth and belief. In particular, the notion of direction of fit is best understood as different ideals, or formal ends, guiding the inference, from what I call “prima-facie” attitudes to what I call “all-out” attitudes respectively in the theoretical and practical realm. However, I’ll argue that there’s no non-circular way of making this distinction. But even if no definition of “belief” and “desire” come out of the distinction between directions of fit, it does elucidate the different natures of practical and theoretical enquiry. However, understood this way, the notion of direction of fit does not seem to capture the distinction illustrated by Anscombe’s example. I try to argue in the last section that Anscombe’s compelling example is best explained not by a distinction between directions of fit, but by a distinction between two different inferential mistakes: one from general to general or particular to general, and the other from general to particular. There’s an important asymmetry between practical and theoretical endeavours in this neighbourhood. However, noticing this asymmetry will also fail to deliver the philosophical payoffs that the notion of direction of fit was supposed to have. (shrink)

Most contemporary action theorists accept – or at least find plausible – a belief condition on intention and a knowledge condition on intentional action. The belief condition says that I can only intend to ɸ if I believe that I will ɸ or am ɸ-ing, and the knowledge condition says that I am only intentionally ɸ-ing if I know that I am ɸ-ing. The belief condition in intention and the knowledge condition in action go hand in hand. After all, if (...) intending implies belief, and if ɸ-ing intentionally implies intending to ɸ, then in ɸ-ing, I intend to be ɸ-ing, and, by the belief condition, I believe that I am ɸ-ing, and if this belief is justified, and we are not in a Gettier situation, etc., then, I will also satisfy the knowledge condition. Moreover, the claim that when intentions properly result in action, the corresponding belief constitutes knowledge is a relatively safe assumption, at least as an assumption about what it is generally the case. (shrink)

Kant’s views on the relation between freedom and moral law seem to undergo a major, unannounced shift. In the third section of the Groundwork, Kant seems to be using the fact that we must act under the idea of freedom as a foundation for the moral law. However, in the Critique of Practical Reason, Kant claims that our awareness of our freedom depends on our awareness of the moral law. I argue that the apparent conflict between the two texts depends (...) on a reading of the opening paragraphs of Groundwork III, and on an interpretation of Kant’s claim that we “act under the idea of freedom”, that is implausible on textual and on philosophical grounds. I then present an alternative interpretation of what Kant means by “acting under the idea of freedom” and of the opening paragraphs of Groundwork III. I argue that the only substantive conclusion of these paragraphs is that no theoretical proof of freedom is necessary. Moreover I argue that although these paragraphs raise concerns about the validity of the moral law, these concerns and Kant’s answers to them, do not give rise to any significant conflict with his views in the Critique of Practical Reason. (shrink)

Collaboration requires agents to coordinate their behavior on the fly, sometimes cooperating to solve a single task together and other times dividing it up into sub‐tasks to work on in parallel. Underlying the human ability to collaborate is theory‐of‐mind (ToM), the ability to infer the hidden mental states that drive others to act. Here, we develop Bayesian Delegation, a decentralized multi‐agent learning mechanism with these abilities. Bayesian Delegation enables agents to rapidly infer the hidden intentions of others by inverse planning. (...) We test Bayesian Delegation in a suite of multi‐agent Markov decision processes inspired by cooking problems. On these tasks, agents with Bayesian Delegation coordinate both their high‐level plans (e.g., what sub‐task they should work on) and their low‐level actions (e.g., avoiding getting in each other's way). When matched with partners that act using the same algorithm, Bayesian Delegation outperforms alternatives. Bayesian Delegation is also a capable ad hoc collaborator and successfully coordinates with other agent types even in the absence of prior experience. Finally, in a behavioral experiment, we show that Bayesian Delegation makes inferences similar to human observers about the intent of others. Together, these results argue for the centrality of ToM for successful decentralized multi‐agent collaboration. (shrink)