Moral Machines?

Wendell Wallach and Colin Allen??s Moral Machines: Teaching Robots Right From Wrong (Oxford University Press, 2009) explores efforts to develop machines that, not only can be employed for good or bad ends, but which themselves can be held morally accountable for what they do??artificial moral agents (AMAs). This essay is a critical response to Wallach and Allen??s conjectures. Although Wallach and Allen do not suggest that we are close to being able to create full-fledged AMAs, they do talk seriously about making incremental progress in the direction of creating them (even if we never fully succeed). However, there are important questions about the moral development of AMAs that Moral Machines does not address. Given the responsibilities entrusted to human moral agents, we take questions about their moral development very seriously. In the case of children, the hope is that eventually they will develop into full-fledged moral agents. How might we expect this to go with less than fully formed AMAs? Will there be a comparable story of moral development and moral education that we can tell?     

Are Machines Gender Neutral? Gender-Stereotypic Responses to Computers With Voices

This study tested whether computers embedded with the most minimal gender cues will evoke gender-based stereotypic responses. Using an experimental paradigm (N = 40) that involved computers with voice output, the study tested 3 gender-based stereotypes under conditions in which all suggestions of gender were removed, with the sole exception of vocal cues. In all 3 cases, gender-stereotypic responses were obtained. Because the experimental manipulation involved no deception regarding the source of the voices. this study presents evidence that the tendency to gender stereotype is extremely powerful, extending even to stereotyping of machines.     

Are multi‐layer backpropagation networks catastrophically amnesic?

Connectionist models with a backpropagation learning rule are known to have a serious problem. Such models exhibit catastrophic interference (or forgetting) with sequential training. Having learned a set of patterns, if the model is trained on another set of patterns, its performance on the first set can dramatically deteriorate very rapidly. The present study reconsiders this issue with four simulations. The model learned arithmetic facts sequentially, but the interference was only modest with random (hence approximately orthogonal) inputs. Essentially the same result was obtained when the inputs are made less orthogonal by adding irrelevant elements. Reducing the number of hidden units did not have major effects. This study suggests that the interference problem has been somewhat overstated.     

Is the human mind a Turing machine?

In this paper I discuss the topics of mechanism and algorithmicity. I emphasise that a characterisation of algorithmicity such as the Turing machine is iterative; and I argue that if the human mind can solve problems that no Turing machine can, the mind must depend on some non-iterative principle — in fact, Cantor's second principle of generation, a principle of the actual infinite rather than the potential infinite of Turing machines. But as there has been theorisation that all physical systems can be represented by Turing machines, I investigate claims that seem to contradict this: specifically, claims that there are noncomputable phenomena. One conclusion I reach is that if it is believed that the human mind is more than a Turing machine, a belief in a kind of Cartesian dualist gulf between the mental and the physical is concomitant.     

Are unsupervised neural networks ignorant? Sizing the effect of environmental distributions on unsupervised learning

Learning environmental biases is a rational behavior: by using prior odds, Bayesian networks rapidly became a benchmark in machine learning. Moreover, a growing body of evidence now suggests that humans are using base rate information. Unsupervised connectionist networks are used in computer science for machine learning and in psychology to model human cognition, but it is unclear whether they are sensitive to prior odds. In this paper, we show that hard competitive learners are unable to use environmental biases while recurrent associative memories use frequency of exemplars and categories independently. Hence, it is concluded that recurrent associative memories are more useful than hard competitive networks to model human cognition and have a higher potential in machine learning.     

Are abilities dispositions?

Synthese - Abilities are in many ways central to what being an agent means, and they are appealed to in philosophical accounts of a great many different phenomena. It is often assumed that...     

Are shapes intrinsic?

It is widely believed that shapes are intrinsic properties. But this claim is hard to defend. I survey all known theories of shape properties, and argue that each theory is either incompatible with the claim that shapes are intrinsic, or can be shown to be false.     

Are Lesbians Women?

I argue that Mordque Wittig's view that lesbians are not women neglects the complexities involved in the composition of the category “woman.” I develop an articulation of the concept “woman” in the contemporary United States, with thirteen distinct defining characteristics, none of which are necessary nor sufficient. I argue that Wittig's emphasis on the material production of “woman” through the political regime of heterosexuality, however, is enormously fruitful for feminist and queer strategizing.     

Are Counterpossibles Epistemic?

It has been suggested that intuitions supporting the nonvacuity of counterpossibles can be explained by distinguishing an epistemic and a metaphysical reading of counterfactuals. Such an explanation must answer why we tend to neglect the distinction of the two readings. By way of an answer, I offer a generalized pattern for explaining nonvacuity intuitions by a stand‐and‐fall relationship to certain indicative conditionals. Then, I present reasons for doubting the proposal: nonvacuists can use the epistemic reading to turn the table against vacuists, telling apart significant from spurious intuitions. Moreover, our intuitions tend to survive even if we clear‐headedly intend a metaphysical reading.     

Are affordances normative?

Phenomenology and the Cognitive Sciences - In this paper we explore in what sense we can claim that affordances, the objects of perception for ecological psychology, are related to normativity....     

Are models necessary?

Modelling is currently one of the most successful and popular ways of dealing with some neurotic problems. It is based on a social learning theory that amongst other virtues takes into account cognitive variables. helps shed light on the aetiology of some neurotic disorders (Rachman. 1972). and seems particularly relevant to the understanding of much child behaviour.The technique of the modelling treatment has been refined over the years, largely through the work of Bandura, and the contention to date has been that behavioural improvement is brought about by vicarious processes (Bandura, 1969). However, since a modelling procedure by necessity involves exposure to a phobic object it is possible that vicarious processes are in fact redundant and that improvement is affected by ‘straightforward’ extinction processes. In short the contention is that interaction with a model disguises the fact that the modelling situation is no more than a modified flooding session. In one of the few studies to test this hypothesis Bandura, Grusec and Menlove (1967) found that exposure alone resulted in some gains for dog-phobic children, but that these changes were neither as great nor as stable as those achieved by similar children in a modelling treatment. Other indirect evidence for this hypothesis can be found by-relating the amount of exposure to the phobic object to outcome in modelling studies. Despite obvious difficulties in cross-trial comparison, rate of success for longer exposure treatments would constitute some support, albeit very tentative, for this hypothesis. Unfortunately, few of the studies contained enough reliable information on the relevant variables to reveal much. Bandura et al. (1967) used 80 min of exposure in all, with success rates of 70%, whilst less exposure time resulted in worse outcome in studies by Spiegler et al. (1969) and Ritter (1969). However, against this, Hill et al. (1968) and Rimm and Mahoney (1969) both report good results with brief participant modelling. The fact that participant modelling seems to be more beneficial than symbolic modelling, may be regarded as some support for the exposure hypothesis, since presumably exposure in these cases is more intense.However, these studies are largely indirect and one test of the hypothesis requires a comparison between groups one of which is exposed to both phobic object and model interaction (symbolic modelling) and one of which is exposed just to the phobic object.     

What Are Quantities?

This paper presents a view of quantities as ‘adverbial’ entities of a certain kind—more specifically, determinate ways, or modes, of having length, mass, speed, and the like. In doing so, it will be argued that quantities as such should be distinguished from quantitative properties or relations, and are not universals but are particulars, although they are not (Fregean) objects, either. A main advantage of the adverbial view over its rivals will be found in its superior explanatory power with respect to both certain fundamental principles of quantity and ordinary quantitative reasoning involving quantitative relations like three times as long as and 2 metres longer than.     

Are adjustments insufficient?

Many judgmental biases are thought to be the product of insufficient adjustment from an initial anchor value. Nearly all existing evidence of insufficient adjustment, however, comes from an experimental paradigm that evidence indicates does not involve adjustment at all. In this article, the authors first provide further evidence that some kinds of anchors (those that are self-generated and known to be incorrect but close to the correct answer) activate processes of adjustment, whereas others (uncertain anchors provided by an external source) do not. It is then shown that adjustment from self-generated anchors does indeed tend to be insufficient, both by comparing the estimates of participants starting from different anchor values and by comparing estimates with actual answers. Thus, evidence is provided of adjustment-based anchoring effects similar to the accessibility-based anchoring effects observed in the traditional anchoring paradigm, supporting theories of social judgment that rely on mechanisms of insufficient adjustment.     

Are beliefs signals?

Eric Funkhouser argues that beliefs can function as social signals. I argue that Funkhouser’s argument for this conclusion rests on a problematic definition of “signal,” and that on standard definitions, the imperceptibility of beliefs disqualifies them from counting as signals. However, I also argue that Funkhouser’s insights about the social functions of beliefs can be true even if his claim that beliefs are signals is false.