A competence framework for artificial intelligence research

ABSTRACT

While over the last few decades AI research has largely focused on building tools and applications, recent technological developments have prompted a resurgence of interest in building a genuinely intelligent artificial agent – one that has a mind in the same sense that humans and animals do. In this paper, I offer a theoretical and methodological framework for this project of investigating “artificial minded intelligence” (AMI) that can help to unify existing approaches and provide new avenues for research. I first outline three desiderata that a framework for AMI research should satisfy. In Section 1, I further motivate the desiderata as well as the need for a new framework. Section 2 develops a general methodological approach, the “generative methodology,” and Section 3 develops a version of this methodology, the “Competence Framework for AI research” (CFAI).

KEYWORDS:

• Artificial intelligence
• competence
• reduction
• computation
• robotics
• mind
• action
• emergence
• embodiment

Acknowledgments

Thanks to Ben Baker, Jared Culbertson, Devin Curry, Louise Daoust, Carolina Flores, Daniel Guralnik, Paul Gustafson, Ting Fung Ho, Daniel Koditschek, Paul Reverdy, Sonia Roberts, Jordan Taylor, Eugene Vaynberg, Mingjun Zhang, the participants in my Fall 2016 and Spring 2018 seminars on artificial intelligence, and the members of the MIRA group at the University of Pennsylvania.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. See also Minsky (1986).

2. Recently, Michael Rescorla (2014) argues that if we assume a counterfactual interventionist account of causation such as that of Woodward (2003), then the content of certain computational states will be causally efficacious. This sort of argument may suggest to the reader that accounts assuming mental states are identical to computational states as part of their methodological framework have more adaptive resources than I have attributed to them. While I do not think that Rescorla’s argument succeeds (and explain why in Miracchi), if it did, it would not provide the computationalist with the kind of flexibility that I am urging we need. Rescorla maintains—as I think anyone endorsing computationalism must—that the causal roles of mental states can be sufficiently described syntactically, even if they can also be described semantically. (He thinks this is a kind of non-problematic overdetermination because of the metaphysical relationships involved.) This view still rules out the possibility that mental kinds are distinctive in the sense that I have articulated, because, according to it, mental causal generalizations can be given in syntactical terms even if they can also be given in semantic terms.

3. Thanks to Michael Strevens for suggesting this example. Please also see Miracchi (2017a) for discussion. Interestingly, Peacocke (1994) also uses this example in order to motivate the idea that mental explanations are explanatorily distinctive. His main focus in the paper, however, is to motivate a conception of computation that is inherently content-involving, and thus to save the computational cognitive science as a fruitful research program from being seen as implausible due to its having an overly reductionist conception of computation. Although it is beyond the scope of this paper to evaluate Peacocke’s account of computation (I discuss this in work in progress), two points are in order. First, Peacocke does think that the classical conception of computation is not content-involving and that, for this reason, mental processes – which are typically ineliminably content-involving—cannot be identified with such computational processes. It is this classical notion of computation that is relevant to our discussion because it is the one that is still employed by AI research in the construction of artificial systems. The (broadly) syntactic account of computation is crucial because its explanatory import in AI research is deeply dependent on the connection between computational and algorithmic characterizations and straightforward implementations in physical systems. Peacocke’s content-involving “computations” are not readily implementable in physical systems because they use unreduced externalist semantic vocabulary. Second, he is optimistic about the prospects of providing a theory that shows how syntactic computation, in the right context, determines contentful computation. While I think that there are reasons not to conceptualize content-involving generalizations as computations, I am happy, for the purposes of this paper, if those more amenable to the idea interpret the proposal I develop here as showing how this project might be rigorously undertaken.

4. See, for example, Fodor (1987), p. xii.

5. Thanks to an anonymous referee for pressing me on this question.

6. Note that the definitional methodology does not require that a single node be identified with the mental state or event. Connectionist accounts typically identify mental states with distributed representations, but these distributed representational outputs are still identified with our explananda without further comment.

7. Frances Egan (2014) adeptly argues that contemporary computational research programs do not have rigorous methods for attributing mental semantic content to computational states. Instead, she argues that the attribution of such content is an informal gloss on these theories and that the only content attributions that are a part of the rigorous theories themselves are the mathematical contents, which are useful because they explain how the mechanisms function at a high level of abstraction, independently of other features of the cognitive system or environment. I think Egan is correct on both accounts; however, I am more optimistic about the possibility of empirical explanations of semantic content. See also Miracchi (2017b) for discussion.

8. This is true of both the cited paper Marr and Poggio (1976) and his monograph Vision (1982).

9. Someone who both endorses computationalism about the mind and accepts the extended-mind thesis (e.g., Clark (2008)) might defend the claim that watching for predators is a computation. However, this kind of view would need defense; it should not fall out of our uninterrogated methodological commitments regarding AI research.

10. Thanks to an anonymous referee for raising this question.

11. Toward the end of the paper, Hassabis et al. (2017) do discuss how AI has benefited and might still benefit neuroscience, but their examples focus on algorithmic and implementational influence rather than on the evaluation of more general hypotheses relating computational characterizations to mental ones.

12. Hassabis et al. (2017) do not draw on this theory in their section on that topic (although some of this work is cited briefly in discussions of imagination and planning).

13. This is a point that Krakauer, Ghazanfar, Gomez-Marin, MacIver, and Poeppel (2017) have recently made within the context of neuroscience, although they too do not cleanly separate behavioral characterizations from computational or algorithmic characterizations as carefully as one should. They characterize behavior as properties of whole organisms, using examples such as bird flight and the modulation of oscillations in electric fish to argue that, in order to understand what the brain is doing, we need detailed descriptions of behavior and these will not be at the same level as the underlying explanatory features.

14. Note that the argument just given does not rely on convincing the reader that the Hassabis- or Maguire-style account of episodic memory is incorrect but only that there may be competing hypotheses that are worth investigating and that an approach that seeks to identify mental and computational processes as part of its methodological framework for AI will be unable to critically evaluate them.

15. Once we clearly distinguish the agent level from the computational and algorithmic levels, there is no longer a sharp divide between how we treat mental processes, such as perception, learning, and memory, and overt actions, such as writing at one’s computer, foraging, and flight. Since none of these is taken to be identical to any computational process at the outset of the investigation, we can remain open about the extent to which and the manner in which the body and the environment play a constitutive role in all of them, and we can revisit assumptions about appropriate computational descriptions of underlying information-processing. An embodied cognition approach naturally falls out of distinguishing the agent and computational levels.

16. Similar approaches can be found in some work in philosophy of the special sciences: Godrey-Smith (2008); Bickle (2008); Bickle and Silva (2009); Woodward (2008); Strevens (2004). See also Klein (2017).

17. This has no relation to generative models as described by Hassabis et al. (2017).

18. Pearl (2000); Halpern and Pearl (2001a, 2001b).

19. Some of my work in progress investigates this idea.

20. There are subtleties here. See Miracchi (2017a) for discussion of cases where neuroscientists manipulate mental features to detect neural features. I argue against Craver (2007) that these should be interpreted differently in order to preserve the asymmetry of the generative relation.

21. See also Smolensky (1991) on two senses of “implementation.”

22. In such cases, it may be inappropriate to call the explanandum model the agent model because agent models are supposed to represent genuine intelligence and agency. We may call the more general class of explanandum models emergent models. Thanks to Daniel E. Koditschek for suggesting this terminology.

23. See, for example, Raghu, Gilmer, Yosinski, and Sohl-Dickstein (2017) for current work on analyzing DNNs.

24. David Chalmers (2011) argues that every system implements a computation because he thinks that a computation is just an abstract causal specification of the system. While I think his account is overly permissive, here I am trying to stay as neutral as possible on which conception of computation is correct. Let it suffice to note that Chalmers’ conception is syntactic, not semantic. If we are interested in explaining mental processes which are characterized semantically, we can still make the relevant distinction between simulation and duplication. Systems that duplicate computational processes in Chalmers’ sense may only simulate the ones we are interested in because they merely instantiate abstract structural relations, not content-involving relations, between mental states. Moreover, he also notes that his account will not apply to relational mental events or processes, such as perceiving an object or watching for predators. So, insofar as those are our explananda, Chalmers’ account will not entail that implementations of formal models of these duplicate such processes.

25. The “what it’s like” locution is due to Nagel (1974).

26. See also Miracchi (2017a) for arguments for the applicability of the generative difference-making framework to explanations of consciousness in cognitive science.

27. This claim, of course, echoes Gibson (1979), although instead of focusing on the agential character of our perceptual experience, I am focusing on the perspectival character of action.

28. Here we can distinguish between what Tyler Burge (2010) calls psychological agency as opposed to biological agency. See pages 327–337. See also Miracchi (2017b) for discussion.

29. What counts as “sufficiently large” often varies with the competence in question. It may be indeterminate or vary with context. Moreover, note that a reliability threshold for competence possession is a metaphysical requirement only; it does not imply anything about how we know whether a person possesses a competence. We may be able to observe directly that the behavior is competent.

30. See, e.g., Miracchi (2015, 2017b, forthcoming).

31. Indeed, the approach to competences I develop in Miracchi (2017b, 2015) is primarily concerned with addressing such problems.

32. See Hawley (2003) and Setiya (2012) for discussion.

33. This is especially true of Jerry Fodor’s work, e.g., Fodor (1987).  See Miracchi (2017b) for more detailed discussion.

34. One criticism of appealing to reliability – and, by extension, competences – in explaining intentionality and purposeful action is a concern that, often, animals are not successful in pursuing their aims (see, e.g., Allen and Bekoff, p. 70). This objection, I think, is mostly due to a confusion about what kind of aims we are talking about. A bird might be motivated to sing but may not have any awareness of what that singing is for, evolutionarily speaking (attracting a mate, warning others of predators, etc.). On my view, the singing, at least, is purposeful, even if attracting a mate might not be. This allows us to see how the mentality of certain less sophisticated animals might be genuine, that is, might involve purposes that the animal has, but it need not involve the kind of sophistication that an evolutionary teleological account might attribute. So, rather than the appeal to reliability being a weakness of the theory, I think it actually helps us to correctly understand the distinction between mental teleology – the possession of purposes by an agent – and other forms of teleology.

35. Note that this does not rule out the possibility that consciousness might require certain kinds of cortical buildup.

36. Schurger and Uithol (2015) build on these results and adduce some others to argue explicitly that “actions emerge from a causal web in the brain, rather than a central origin of intentional action” (761).

37. Russell and Norvig define autonomy thus: “A rational agent should be autonomous – it should learn what it can to compensate for partial or incorrect prior knowledge. After sufficient experience of its environment, the behavior of a rational agent can become effectively independent of its prior knowledge. Hence, the incorporation of learning allows one to design a single rational agent that will succeed in a vast variety of environments” (p. 39). See also Bekey (2005).

38. A fair coin is such that Pr(coin lands heads coin flipped) = 0.5. I am here rejecting a frequency account of probability in favor of objective probabilities, in particular, objective conditional probabilities. See Ha´Jek (2007) for discussion.

39. Many thanks to Daniel Koditschek for his comments on this portion of the paper.

40. More precisely, we can define a Lyapunov function as follows (Hirsch, Smale, and Devaney (2004), p. 193). Let $X\ast$ be an equilibrium point for X' = F(X). Let $\mathrm{L}:\mathrm{O}\to \mathrm{R}$ be a differentiable function defined by an open set O containing $X\ast$. Let (1) L($X\ast$) = 0 and L(X) > 0 if $X\ne X\ast$ and (2) L^. ≤ 0 in O − $X\ast$. Then L is a Lyapunov function for $X\ast$.

41. One issue requiring care, but amenable to technical resolution, would be to ensure that the domain of our (presumably continuously differentiable) Lyapunov function is itself continuously differentiable. This will likely be non-trivial for other than spatiotemporal dimensions, requiring either significant idealization or very relaxed systems models. Still, various idealized “smoothings” of the space in order to make it exhibit the necessary formal properties (A. M. Johnson & Koditschek, 2016) or weakening of the presumed transition structure (Erdmann, 2010) may help us to bring out key structural features. Moreover, if such a model could not be made to work because of difficulties with developing an adequately behaved dynamics, this would be interesting and valuable information.

42. It provides this kind of structure while nevertheless resisting the temptation to specify the difference between achievements and failures in sub-personal terms from the armchair. For example, in the previous work, I have characterized degenerate exercises of competence as cases where the sub-personal cognitive basis is the same as it would be in a case of achievement, but something about the body or environment is different (Miracchi, 2017b, 2015). Although this way of specifying degenerate exercises may be extensionally adequate, it might not always get to the heart of what makes a degenerate case degenerate.

Additional information

Notes on contributors

Lisa Miracchi

Lisa Miracchi is an Assistant Professor of Philosophy at the University of Pennsylvania.