Over the first year of life, infants gain conceptual skills which allow them to construe semantically related items as similar, even when they have few if any directly-perceived attributes in common. Moreover, this skill first encompasses only broad semantic categories, and only later extends to more subtle distinctions, when conceptual and perceptual similarity relations do not coincide. In this paper we suggest that a new mechanism must be added to the mix of possible bases for this observed developmental change. In agreement with many others, we suggest that infants’ earliest conceptual representations are organised with respect to certain especially useful or salient properties, regardless of whether such properties can be directly observed. However we suggest that in many cases this salience may itself be acquired, through domain-general learning mechanisms that are sensitive to the high-order coherent covariation of directly-observed stimulus properties across a breadth of experience. To support this argument we will describe simulations with a simple PDP model of semantic memory. When trained with backpropagation to complete queries about the properties of different objects, the model’s internal representations differentiate in a coarse-to-fine manner. As a consequence, different sets of properties come to be especially “salient” to the

Autonomous agents make frequent use of knowledge in the form of categories --- categories of objects, human gestures, web pages, and so on. This paper describes a way for agents to learn such categories for themselves through interaction with the environment. In particular, the learning algorithm transforms raw sensor readings into clusters of time series that have predictive value to the agent. We address several issues related to the use of an uninterpreted sensory apparatus and show specific examples where a Pioneer 1 mobile robot interacts with objects in a cluttered laboratory setting.

Recent developments in philosophy, linguistics, developmental psychology and artificial intelligence make it possible to envision a developmental path for an artificial agent, grounded in activity-based sensorimotor representations. This paper describes how Neo, an artificial agent, learns concepts by interacting with its simulated environment. Relatively little prior structure is required to learn fairly accurate representations of objects, activities, locations and other aspects of Neo's experience. We show how classes (categories) can be abstracted from these representations, and discuss how our representation might be extended to express physical schemas, general, domain-independent activities that could be the building blocks of concept formation.

Biological organisms display an astonishing capability to learn new skills and adapt to dynamic environments that far outperforms any computer or robot system. This paper presents an approach to robot skill acquisition that takes concepts from developmental theory to structure the learning problem and provides a mechanism to generate developmental schedules for a robot systems. The approach uses a developmental assembler to construct reusable and temporally extended actions in a sequence. All behavior is initially constructed from a set of innate control laws and events that delineate control decisions are derived from the pattern of (dis)equilibria on a working subset of sensorimotor policies. We show how this architecture can be used to accomplish sequential knowledge gathering and representation tasks and provide examples of developmental learning using a quadrupedal walking robot.