Focusing Attention for Observational Learning: The Importance of Context

A significant component of human observational learning is the ability to focus attention toward important or relevant input features. Amechanism with this capability can serve as an inductive bias to facilitate learning in both humans and machines. Past attempts to model attentional focus for human learning have postulated a single salience value for each feature, such that features with greater salience command more attention. These models, however, assume that the feature's salience is not dependent on context, whereas studies of human attention show sensitivity to context. This paper presents a mechanism for contextually focused attention in observational learning. 1 Introduction Observational learning is a form of inductive knowledge acquisition in which there is no external guidance, such as explicit feedback. However, some guidance or learning bias is required to make general induction tractable (eg. Rendell et al., 1987). Since humans do engage in some observational learning (Billman et al., 1987), there must be a method for internally guiding this learning. Discovering such methods will prove useful both for understanding human learning and for designing computer programs that learn from observation (eg. Fisher, 1987). Zeaman and House (19C3) and Billman and licit (1988) have argued that attention directed by learnable feature saliences may provide some internal guidance for human learners. They each proposed a mechanism for doing this and were able to confirm the approach for simple learning. Neither method used context — what already is known about an example — to help focus attention. Other researchers, however, have found that human attention and other cognitive processes vary with context As well, the use of context can allow learning of more *The author wishes to thank Dorrit Billman for providing constructive comments on all aspects of this research, and Janet Kolodner an anonymous reviewer for assistance with earlier versions of the manuscript. complex examples. The noncontext approach assumes that there is only one important subset of features that are always salient. When this assumption is violated, learning is not facilitated. Given that attention is useful for human and machine observational learning, it is important to ensure that proposed attentional mechanisms support a useful type of learning. Human observational learning most clearly is useful for natural language and concept based predictions. Additionally, many machine learning studies of observational learning have concentrated on concept based prediction (Schlimmer, 1986; Fisher, 1987). These types of knowledge have frequently been described as capturing correlational feature structure …