Minimal Nativism: How does cognitive development get off the ground?

Minimal Nativism: How does cognitive development get off the ground? Tomer D. Ullman and Joshua B. Tenenbaum Department of Brain and Cognitive Sciences, MIT Cambridge, MA USA 02139 Noah D. Goodman Department of Psychology, Stanford University Stanford, CA USA 94350 Shimon Ullman Department of Computer Science and Applied Mathematics, Weizmann Institute of Science Rehovot 76100 Israel Elizabeth Spelke Department of Psychology, Harvard University Cambridge, MA USA 02138 Keywords: Cognitive development; Learning; Computational modeling Core knowledge; nurture” debates. At the same time, recent empirical work with young children offers striking new data that both motivates and challenges these computational accounts. Our symposium brings together some of the researchers who have contributed to these developments from both computational and empirical perspectives (Goodman, Ullman, & Tenenbaum, 2011; Spelke & Kinzler, 2006; Tenenbaum, Kemp, Griffiths, & Goodman, 2011; Ullman, Harari, & Dorfman, 2012; Ullman, Goodman & Tenenbaum, 2010; Xu & Kushnir, 2012). Our goals are to survey the landscape of developmental possibilities across multiple domains of physics, psychology, number, geometry, and language; to bring recent models and empirical work into closer contact; and to confront, honestly and clearly, the deep challenges that remain unaddressed. When constructing a mind, what are the basic materials, structures and blueprints a young child has to work with? Are most of the structures already in place, with children merely working to embellish them? Do children begin with several buildings already in place (the Physics Building, the Social Building, the Number building, etc.), and only decorate a bit as they get older, perhaps building bridges between them using language? Such a view might describe a strong innate core hypothesis (Spelke et al., 1994). Or does the child begin with more of an empty plain, and an ability to construct whatever is necessary out of whatever materials are at hand at the time? Such a view might be more along the lines of classic empiricism (Quine, 1964). Our plan is to have four 15-minute talks, followed by a 30-40 minute discussion. T. Ullman will speak first, sketching out the space of potential approaches to a “minimal scaffolding” for cognitive development, and touching briefly on his own work modeling the development of intuitive physics, intuitive psychology, and the interface between these domains. N. Goodman will then present the “probabilistic language of thought” view – that an innate, abstract, domain-general, language-like ability for composing and manipulating conceptual representations is the minimal structure necessary for learning, potentially supplemented with specific 'named-functions' or input- analyzers for certain domains. S. Ullman will then expand on the notion of innate perceptual input analyzers, illustrating with a case study drawn from his recent work on computer vision systems that learn to identify and reason about agents and actions in real-world video. E. Spelke will approach these issues from the standpoint of her recent work on the development of space, number and other mathematical concepts. She will also provide a more general critical perspective on the various computational perspectives presented earlier. This will set the stage for our Many other views are possible, lying somewhere between the extremes of positing that the child starts with everything, and positing that the child starts with nothing. For example, perhaps the child begins with a powerful general-purpose learning mechanism and a general blueprint for how to organize the world’s entities into core domains, but no detailed, specific understanding of how these domains operate. Or perhaps the child begins with a powerful learning mechanism and a general blueprint for cognitive architecture, but no abstract concepts – only raw sensory experience. Yet if her sensory experience can be structured by a few crucial ‘proto-concepts’ - low-level input analyzers that tug her learning apparatus in certain appropriate directions – that minimal scaffolding could be sufficient. Of course metaphors for cognitive development will only take us so far. In the last few years, a number of stimulating proposals for how cognitive development might get off the ground have been framed by computational modeling researchers, and these models offer to bring greater precision, clarity and subtlety to classic “nature versus