A non-Bayesian Account of the "Causal Reasoning" in Sobel, Tenenbaum & Gopnik (2004)
Masanori Nakagawa, Kayo Sakamoto, Asuka Terai · Proceedings of the Annual Meeting of the Cognitive Science Society · 2005
A non-Bayesian Account of the “Causal Reasoning” in Sobel, Tenenbaum & Gopnik (2004) Serban C. Musca ([email protected]) Psychology and NeuroCognition Lab - CNRS UMR 5105, University Pierre Mendes France (Grenoble 2) 1251 Avenue Centrale, BP 47, 38040 Grenoble cedex 9, France Gautam Vallabha ([email protected]) Center for the Neural Basis of Cognition, Carnegie Mellon University, 4400 Forbes Avenue Pittsburgh, PA 15213, USA Abstract Sobel, Tenenbaum & Gopnik (2004) investigated the development of causal inferences in preschoolers in three experiments with tasks adapted from conditioning literature (backwards blocking and screening-off) and concluded from this indirect evidence that children develop a mechanism for Bayesian structure learning. It is proposed that (a) the differential performances in the two tasks are more likely due to differential memory demands, and (b) the observed developmental differences between 3½ and 4½-year old children may be due to maturation of the memory system, with higher retroactive interference in younger children and lower retroactive interference in older children. This account is supported by simulations with Ans & Rousset's (1997, 2000) memory self-refreshing neural networks architecture. The implications of the account proposed here on a theory of causal relation learning are discussed. Keywords: causal inference; retroactive interference; backwards blocking; screening-off; memory limitations; preschoolers; developmental maturation; memory self-refreshing; artificial neural networks. Introduction Early knowledge of the causal structure of the world is thought to result from innate abilities or from interactions with the environment during early childhood. In the latter framework, a learning mechanism must be specified in order to delineate a theory of causal relation learning. Sobel, Tenenbaum & Gopnik (2004) have recently proposed such a theory, suggesting that children construct a ‘causal graph’ – an abstract representation of the causal structure of a set of variables – based on evidence about the conditional probability of those variables (Sobel et al., 2004, p. 306). In particular, they proposed that children use Bayesian reasoning to construct the causal graph, and tested these claims in two experiments. In both, children were told that only certain objects (called blickets) cause a device (a blicket-detector) to be activated. In the “indirect screening- off” task, the children were shown that the detector is activated when two objects (A and B) are placed on it, and that it does not activate when object A is placed on it by itself. Then, they were asked if each object was a blicket. In the “backwards blocking” task, the detector is activated when two objects (A and B) are placed on it, and also when object A is placed on it by itself. Sobel et al. found that 4½-year old children, and to a lesser extent, 3½-year old children, were both able to make the expected inferences, that is that object B is a blicket in the indirect screening-off task, but not in the backwards-blocking task. Sobel et al. (2004) used these results to argue that children’s responses are based on Bayesian structure learning rather than on learning of cause-consequence associations. Though attractive and nicely formalized, the Bayesian account has two problematic limitations. First, there is a conceptual problem. A Bayesian inference structure cannot operate without an initial core of knowledge. If the theory aims to explain the origins of this core of knowledge, one is faced with a chicken and egg problem: children are supposed to apply their statistical knowledge in order to enhance some pre-existing knowledge database , but this cannot explain where the initial core of knowledge comes from. Second, the tasks used by Sobel et al. (2004) are adapted from conditioning literature, and are designed to tap into the memory system. To use these tasks as measures of causal reasoning, one has to assume that memory demands are the same for both groups of children. This assumption is erroneous: Both the indirect screening-off and backwards- blocking results may be shown to be memory artifacts rather than instances of causal reasoning. Further, the performance difference between the 4½ and 3½-year-olds can be explained as a maturation of the memory system rather than the development of a Bayesian mechanism. Our critique is based on the simulation of memory as a “self-refreshing neural network”. Memory as a self-refreshing neural network A common problem with neural network models of memory is that of catastrophic forgetting. The memory of a neural network resides in connection weights that are adjusted to improve the network’s performance on the current training set. Consequently, training on a new set S 2 tends to overwrite the effects of prior training on set S 1 (McCloskey & Cohen, 1989; Ratcliff, 1990). The problem can be avoided if sequential learning (i.e. first S 1 then S 2 ) is transformed into concurrent learning (S 1 and S 2 trained together). As concurrency is implausible for sequential learning (e.g. Blackmon et al., in press), it can be