Gaze following: how (not) to derive predictions from a computational model

Abstract

We thank the four commentators for carefully evaluating our model (Triesch, Teuscher, Deak, & Carlson, 2006) and sharing their opinions. We will respond to the commentaries one by one. Chris Moore focuses on two important limitations of our model: our choice not to incorporate attentional cueing mechanisms, and our choice to ignore spatial aspects of gaze following. We agree with his comments. It is important to emphasize, however, that there are good reasons to develop computational models in an incremental fashion. Richardson and Thomas, in their commentary, put it very nicely: ‘Overly complex models are time consuming to build and run the risk of revealing little about the potential causes of a particular behavior, since credit and blame assignment can become opaque.’ Our choice was thus to start with a very simple model, which we view as a useful stepping stone for the development of more complex and powerful models. In fact, our current work has been extending the present model in the suggested directions (see Jasso, Triesch & Teuscher, 2005, for recent results). Gergely Csibra’s commentary nicely illustrates some pitfalls of deriving and interpreting the predictions of a computational model. In a first manipulation of our model, Csibra fixes the only rewarding visual stimulus (apart from the caregiver) to one location. In this situation, there are only two interesting things to see in the environment and they are always in the same place. Clearly, the model infant ‘growing up’ in this environment has nothing to gain by learning to follow gaze because the caregiver provides no additional information regarding where the other interesting thing is. In fact, whenever the caregiver looks away from the fixed target, it is guaranteed that there will be nothing interesting to see there. Not surprisingly, the model infant does not learn to follow gaze, but learns to largely ignore the caregiver’s looking behavior and look directly to the fixed target. Thus, the model behavior is easily explained and understood given the assumptions made. But what conclusions should be drawn from this finding? Csibra thinks it means that our model predicts that ‘Infants raised in stable environments will be slower, or less likely, to develop a gaze following response than infants raised in unpredictably changing environments.’ What he ignores is that his manipulation also made the caregiver’s gaze predict the absence of a rewarding sight in all but one location, violating a fundamental assumption of our model. Thus, Csibra’s interpretation is misleading. His second prediction is somewhat better justified. Based on our experiments with multiple targets he suggests that: ‘The richer the environment in which an infant grows up, the less likely (or slower) that she will learn to follow gaze.’ This statement is essentially correct, but it ignores the fact that in an impoverished environment where there is nothing interesting to see at certain times, gaze following will also be slow to emerge. Thus, too much or too little ‘richness’ of the environment are both detrimental. We find neither prediction surprising. That too rich an environment may not be optimal for the emergence of gaze following is arguably supported by the observation that most experiments on gaze following find it necessary to use rooms with plain, undistracting walls. Finally, Csibra extends our model (a good idea in principle) and derives two further predictions from his extension to our model. It is very difficult to comment on these predictions, simply because Csibra’s description