Capturing relational capacity

We agree with Lake et al. that structured relational representations are essential for human cognition. But that raises the question of how such representations are acquired and used. There is abundant evidence from both children and adults that structure mapping (Gentner Reference Gentner1983) is a major route to acquiring and using knowledge. For example, physicists asked to solve a novel problem spontaneously use analogies to known systems (Clement 1988), and studies of working microbiology laboratories reveal that frequent use of analogies is a major determinant of success (Dunbar Reference Dunbar, Sternberg and Davidson1995). In this respect, children are indeed like little scientists. Analogical processes support children's learning of physical science (Chen & Klahr Reference Chen and Klahr1999; Gentner et al. 2016) and mathematics (Carey Reference Carey2009; Mix Reference Mix1999; Richland & Simms Reference Richland and Simms2015). Analogy processes pervade everyday reasoning as well. People frequently draw inferences from analogous situations, sometimes without awareness of doing so (Day & Gentner Reference Day and Gentner2007).

Moreover, computational models of structure mapping's matching, retrieval, and generalization operations have been used to simulate a wide range of phenomena, including geometric analogies, transfer learning during problem solving, and moral decision making (Forbus et al. Reference Forbus, Ferguson, Lovett and Gentner2017). Simulating humans on these tasks requires between 10 and 100 relations per example. This is a significant gap. Current distributed representations have difficulty handling even one or two relations.

Even visual tasks, such as character recognition, are more compactly represented by a network of relationships and objects than by an array of pixels, which is why human visual systems compute edges (Marr Reference Marr1983; Palmer Reference Palmer1999). Further, the results from adversarial training indicate that deep learning systems do not construct human-like intermediate representations (Goodfellow et al. Reference Goodfellow, Schlens and Szegedy2015; see also target article). In contrast, there is evidence that a structured representation approach can provide human-like visual processing. For example, a model that combines analogy with visual processing of relational representations has achieved human-level performance on Raven's Progressive Matrices test (Lovett & Forbus Reference Lovett and Forbus2017). Using analogy over relational representations may be a superior approach even for benchmark machine learning tasks. For example, on the link plausibility task, in which simple knowledge bases (Freebase, WordNet) are analyzed so that the plausibility of new queries can be estimated (e.g., Is Barack Obama Kenyan?), a combination of analogy and structured logistic regression achieved state-of-the-art performance, with orders of magnitude fewer training examples than distributed representation systems (Liang & Forbus Reference Liang and Forbus2015). Because structure mapping allows the use of relational representations, the system also provided explanations, the lack of which is a significant drawback of distributed representations.

Causality and qualitative models

Lake et al. focus on Bayesian techniques and Monte Carlo simulation as their alternative explanation for how human cognition works. We agree that statistics are important, but they are insufficient. Specifically, we argue that analogy provides exactly the sort of rapid learning and reasoning that human cognition exhibits. Analogy provides a means of transferring prior knowledge. For example, the Companion cognitive architecture can use rich relational representations and analogy to perform distant transfer. Learning games with a previously learned analogous game led to more rapid learning than learning without such an analog (Hinrichs & Forbus Reference Hinrichs and Forbus2011). This and many other experiments suggest that analogy not only can explain human transfer learning, but also can provide new techniques for machine learning.

Our second major claim is that qualitative representations – not quantitative simulations – provide much of the material of our conceptual structure, especially for reasoning about causality (Forbus & Gentner Reference Forbus and Gentner1997). Human intuitive knowledge concerns relationships such as “the higher the heat, the quicker the water will boil,” not the equations of heat flow. Qualitative representations provide symbolic, relational representations of continuous properties and an account of causality organized around processes of change. They enable commonsense inferences to be made with little information, using qualitative mathematics. Decades of successful models have been built for many aspects of intuitive physics, and such models have also been used to ground scientific and engineering reasoning (Forbus Reference Forbus2011). Moreover, qualitative models can explain aspects of social reasoning, including blame assignment (Tomai & Forbus Reference Tomai and Forbus2008) and moral decision making (Dehghani et al. Reference Dehghani, Tomai, Forbus and Klenk2008), suggesting that they are important in intuitive psychology as well.

We note two lines of qualitative reasoning results that are particularly challenging for simulation-based accounts. First, qualitative representations provide a natural way to express some aspects of natural language semantics, for example, “temperature depends on heat” (McFate & Forbus Reference McFate, Forbus, Papafragou, Grodner, Mirman and Trueswell2016). This has enabled Companions to learn causal models via reading natural language texts, thereby improving their performance in a complex strategy game (McFate et al. Reference McFate, Forbus and Hinrichs2014). Second, qualitative representations combined with analogy been used to model aspects of conceptual change. For example, using a series of sketches to depict motion, a Companion learns intuitive models of force. Further, it progresses from simple to complex models in an order that corresponds to the order found in children (Friedman et al. Reference Friedman and Forbus2010). It is hard to see how a Monte Carlo simulation approach would capture either the semantics of language about processes or the findings of the conceptual change literature.

Although we differ from Lake et al. in our view of intuitive physics and the role of analogical processing, we agree that rapid computation over structured representations is a major feature of human cognition. Today's deep learning systems are interesting for certain applications, but we doubt that they are on a direct path to understanding human cognition.