We were encouraged by the broad enthusiasm for building machines that learn and think in more human-like ways. Many commentators saw our set of key ingredients as helpful, but there was disagreement regarding the origin and structure of those ingredients. Our response covers three main dimensions of this disagreement: nature versus nurture, coherent theories versus theory fragments, and symbolic versus sub-symbolic representations. These dimensions align with classic debates in artificial intelligence and cognitive science, although, rather than embracing these debates, we emphasize ways of moving beyond them. Several commentators saw our set of key ingredients as incomplete and offered a wide range of additions. We agree that these additional ingredients are important in the long run and discuss prospects for incorporating them. Finally, we consider some of the ethical questions raised regarding the research program as a whole.

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 (1) build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems; (2) ground learning in intuitive theories of physics and psychology to support and enrich the knowledge that is learned; and (3) 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.

If Bayesian Fundamentalism existed, Jones & Love's (J&L's) arguments would provide a necessary corrective. But it does not. Bayesian cognitive science is deeply concerned with characterizing algorithms and representations, and, ultimately, implementations in neural circuits; it pays close attention to environmental structure and the constraints of behavioral data, when available; and it rigorously compares multiple models, both within and across papers. J&L's recommendation of Bayesian Enlightenment corresponds to past, present, and, we hope, future practice in Bayesian cognitive science.

We present a hierarchical Bayesian framework for modeling the acquisition of verb argument constructions. It embodies a domain-general approach to learning higher-level knowledge in the form of inductive constraints (or overhypotheses), and has been used to explain other aspects of language development such as the shape bias in learning object names. Here, we demonstrate that the same model captures several phenomena in the acquisition of verb constructions. Our model, like adults in a series of artificial language learning experiments, makes inferences about the distributional statistics of verbs on several levels of abstraction simultaneously. It also produces the qualitative learning patterns displayed by children over the time course of acquisition. These results suggest that the patterns of generalization observed in both children and adults could emerge from basic assumptions about the nature of learning. They also provide an example of a broad class of computational approaches that can resolve Baker's Paradox.

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.

Rogers & McClelland (R&M) criticize models that rely on structured representations such as categories, taxonomic hierarchies, and schemata, but we suggest that structured models can account for many of the phenomena that they describe. Structured approaches and parallel distributed processing (PDP) approaches operate at different levels of analysis, and may ultimately be compatible, but structured models seem more likely to offer immediate insight into many of the issues that R&M discuss.

In order to account for how children can generalize words beyond a very limited set of labeled examples, Bloom's proposal of word learning requires two extensions: a better understanding of the “general learning and memory abilities” involved, and a principled framework for integrating multiple conflicting constraints on word meaning. We propose a framework based on Bayesian statistical inference that meets both of those needs.

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.

We address two kinds of criticisms of our Bayesian framework for generalization: those that question the correctness or the coverage of our analysis, and those that question its intrinsic value. Speaking to the first set, we clarify the origins and scope of our size principle for weighting hypotheses or features, focusing on its potential status as a cognitive universal; outline several variants of our framework to address additional phenomena of generalization raised in the commentaries; and discuss the subtleties of our claims about the relationship between similarity and generalization. Speaking to the second set, we identify the unique contributions that a rational statistical approach to generalization offers over traditional models that focus on mental representation and cognitive processes.

We applaud Millikan's psychologically plausible version of the causal theory of reference. Her proposal offers a significant clarification of the much-debated relation between concepts and beliefs, and suggests positive directions for future empirical studies of conceptual development. However, Millikan's revision of the causal theory may leave us with no generally satisfying account of concept individuation in the mind.