CHILDREN'S CAUSAL INFERENCES AS REVEALED BY BACKWARDS BLOCKING TASKS: A MEMORY SELF-REFRESHING NEURAL NETWORKS ACCOUNT SERBAN C. MUSCA Laboratoire de Psychologie et NeuroCognition — CNRS UMR 5105 Université Pierre Mendès France, BP 47, 38040 Grenoble Cedex 9, France E-mail: Serban.Musca@upmf-grenoble.fr Children aged 3½ exhibit less backwards blocking effect than those aged 4½; the latter only are sensitive to probabilities (Sobel, Tenenbaum & Gopnik, 2004). The original account proposed by Sobel et al. (2004) is that children develop a mechanism for Bayesian structure learning. This account is problematic because it evades the explanation of the origins of the initial core of knowledge that is used by the posited Bayesian mechanism. I propose here an alternative explanation: Children's differential performance stems from a memory limitation, with retroactive interference stronger in younger children, but adult-like in older children. This claim is supported by simulations with Ans and Rousset's (1997) memory self-refreshing neural networks architecture. 1. Introduction Recently, in order to explain how the causal structure of the world is learned from interactions with the environment during early childhood, Sobel, Tenenbaum & Gopnik (2004) proposed that children develop a mechanism for Bayesian structure learning. The key experiments in support of this view used backwards blocking (hereafter BB) tasks and showed that children aged 4½ exhibit more BB than children aged 4½ (experiment 1), and that only the latter are sensitive to a manipulation of the probability of the outcome of interest in a pre-experimental phase (experiment 3). I propose here an alternative explanation. The developmental difference in performance found by Sobel et al. (2004) with BB tasks is a by-product of a developing memory system in children: that of younger children, still immature, leads to a higher level of retroactive interference. Backwards blocking is part of the broader class of retrospective revaluation tasks, where changes in the learned response to a cue (X) arise without further presentations (i.e. training) of that cue but due to the training of an associate of X (e.g., cue Y) during a second phase of training. In the present chapter, I first show that the results of experiment 1 of Escobar et al. (2002) with adults can be simulated with Ans and Rousset's (1997) memory self-refreshing neural networks architecture (hereafter DRSR). In A. Cangelosi, G. Bugmann & R. Borisyuk (Eds.), Progress in Neural Processing (Vol. 16): Modeling Language, Cognition, and Action (pp. 367-371). World Scientific.