Learning in the active mode

Abstract

The paper distinguishes between two different modes of learning by neural networks. Traditional networks learn in the passive mode by incorporating in their internal structure the regularities present in the input and teaching input they passively receive from outside. Networks that live in a physical environment (ecological networks) can learn in the active mode by acting on the environment and learning to predict what changes in the environment or in their relation to the environment are caused by their actions. Being able to predict the consequences of one's own actions is useful when one wants to cause desired consequences with these actions. The paper contrasts learning to predict the consequences of one's actions with learning to predict environmental changes that are independent from the network's actions. It then discusses how perceptually ‘hidden’ properties of the environment such as the weight of objects are better learned in the active rather than in the passive mode and how learning in the active mode can be particularly useful in a social environment and in learning by imitating others. Learning in the active mode appears to be a crucial component of the human adaptive pattern and is tightly linked to another component of this pattern, i.e., the human tendency to modify the external environment rather than adapt to the environment as it is.