At last, a truly developmental account of a fundamental cognitive “skill,” analogy! Leech et al.'s computational demonstration that reasoning by analogy can in principle arise from pattern completion by simple memory processes is extremely important for analogical theory. Recent advances in cognitive neuroscience and connectionism are helping us to understand how the brain builds a complex cognitive system on the basis of simple incremental learning mechanisms that begin with sensory input (Goswami Reference Goswami2008). Indeed, aspects of development long considered to require innate “pre-knowledge,” such as syntax acquisition, can in principle be acquired by networks that learn according to simple statistical algorithms (Freudenthal et al. Reference Freudenthal, Pine and Gobet2006). In developmental psychology, connectionist demonstrations have been crucial both for dealing with the “poverty of the stimulus” argument (Elman Reference Elman2005) and for establishing that relational information can be computed from instance-based learning. Leech et al. achieve the same outcomes for analogy. Simple sensory input (apples, cut apples, knives) can be rich in relational information, and instance-based learning can yield complex cognition.

In my view, the question of whether the brain achieves analogy on the basis of relational priming is secondary to the demonstration that analogies can in principle be achieved by low-level, automatic, and incremental learning processes. Indeed, cognitive neuroscience and developmental psychology both provide extensive empirical demonstrations that relations are represented as more than “transformations between items” (sect. 2.2, para. 2). Instead, low-level sensory processing represents spatio-temporal dynamic structure. In adult neuroscience, it can be shown that when cross-modal input has temporal patterning, the brain will abstract these patterns and alter uni-modal sensory processing in terms of the higher-order dependencies in the patterns (e.g., Noesselt et al. Reference Noesselt, Rieger, Schoenfeld, Kanowski, Hinrichs, Heinze and Driver2007). Regarding babies, 3-month-olds who view abstract motion patterns (point light displays) that specify either vehicles or animals distinguish the two inputs on the basis of this dynamic information alone (and do so as well as they distinguish actual pictures of vehicles and animals; Arterberry & Bornstein Reference Arterberry and Bornstein2001). To babies, relations specify as much information as the objects themselves.

Our sensory processing systems therefore end up prioritising spatio-temporal dynamic structure (abstracted dependencies) over instance-based featural (object) information, and this is also revealed by sensory “illusions” (Riecke et al. Reference Riecke, van Opstal, Goebel and Formisano2007). The primacy given to these abstracted dependencies is very important theoretically, as it means that the way in which our brains process sensory structure can in principle yield the “relational primacy” that I argued for in 1991 (see target article, sect. 3.4, para. 7). In essence, sensory systems are representing underlying structure (traditionally discussed as “prototypes,” “naïve theories,” “innate biases,” or “core knowledge”; e.g., Rosch Reference Rosch, Rosch and Lloyd1978; Spelke Reference Spelke1994). Incremental learning by sensory neural networks that represent dynamic relations automatically represents relational structure. These emergent knowledge systems are enriched and transformed as the child acquires language (Vygotsky Reference Vygotsky1978). In analogy, as in conceptual development, verbal labeling supports structural similarity over perceptual similarity (Gelman & Coley Reference Gelman and Coley1990).

The modelling conducted by Leech et al. is thus compelling in establishing two central developmental phenomena: Analogical completion is an emergent property of the way that relational information is represented in a (neural) network; and incremental learning processes can yield developmental effects previously explained by symbolic theories. Do we need the extra assumptions, that relations are transformations, and that consequently a strong test of the relational priming account is whether semantic relational priming is found in young children (target article, sect. 5.2, para. 4)? I doubt it.

Firstly, relational priming as discussed by Leech et al. and as tested in the studies on adults they cite is rather narrow in scope (e.g., “apple” and “cake” priming “made of”). If children didn't show these automatic effects (which also require adequate reading skills, noise-free reaction times [RTs], and relevant domain knowledge), I am not sure it would matter. Secondly, priming in the cognitive neuroscience literature offers a general tool for studying the nature of the code in a given brain region, via repetition suppression (Dehaene et al. Reference Dehaene, Naccache, Cohen, LeBihan, Mangin, Poline and Rivière2001). This latter conceptualisation of priming seems more relevant to analogy as an emergent phenomenon, and gives inhibition “for free.” Neural repetition suppression techniques have already been used to investigate the coding of relational information (e.g., numerical quantity; see Naccache & Dehaene Reference Naccache and Dehaene2001). Hence, to show that Leech et al's model is biologically plausible, all that is required is evidence that young children show neural repetition suppression to relational information. Even simple causal information (e.g., launching events) would suffice for such a test. Priming effects seem unlikely to be isolated to the temporal cortices (sect. 5.4, para. 7), as semantic memory is no longer understood as a distinct symbolic system. Rather, the activation of particular concepts (by adults) produces neural activation in the sensory modalities associated with those concepts and in association areas recording the conjunctions of particular sets of sensory information (e.g., Barsalou et al. Reference Barsalou, Simmons, Barbey and Wilson2003). Studying the repetition suppression of such conjunctions appears the most productive way to understand analogy as a form of neural priming.

Finally, what of the “unexpected consequence” of the Leech et al. model (sect. 5.1.1, para. 2), that there is no necessary relational shift for any given relation in a child's similarity judgements? This seems highly likely, and a relational primacy account must predict that for some relations, children might show an initial bias for relations over objects. In fact, 3-month-old babies do just this. In the 1980s, Rovee-Collier and her colleagues conducted a series of experiments on infant memory, using a “conjugate reinforcement” paradigm. Three-month-old babies were trained to kick when they saw a distinctive mobile, and were reinforced because their kicking set the mobile in motion (a causal relation). These memories lasted for months, and the most effective retrieval cues were other mobiles, including mobiles that were visually completely different from the distinctive training mobile (e.g., a mobile with one butterfly on a ring as reminder cue, a mobile of hanging dice during relational learning; see Greco et al. Reference Greco, Hayne and Rovee-Collier1990). Thus shared functional relations and not object similarity was the core retrieval cue – consistent with a relational primacy account and with Leech et al.'s model.