Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Computational Theories of Curiosity-Driven Learning (1802.10546v2)

Published 28 Feb 2018 in cs.AI and cs.LG

Abstract: What are the functions of curiosity? What are the mechanisms of curiosity-driven learning? We approach these questions about the living using concepts and tools from machine learning and developmental robotics. We argue that curiosity-driven learning enables organisms to make discoveries to solve complex problems with rare or deceptive rewards. By fostering exploration and discovery of a diversity of behavioural skills, and ignoring these rewards, curiosity can be efficient to bootstrap learning when there is no information, or deceptive information, about local improvement towards these problems. We also explain the key role of curiosity for efficient learning of world models. We review both normative and heuristic computational frameworks used to understand the mechanisms of curiosity in humans, conceptualizing the child as a sense-making organism. These frameworks enable us to discuss the bi-directional causal links between curiosity and learning, and to provide new hypotheses about the fundamental role of curiosity in self-organizing developmental structures through curriculum learning. We present various developmental robotics experiments that study these mechanisms in action, both supporting these hypotheses to understand better curiosity in humans and opening new research avenues in machine learning and artificial intelligence. Finally, we discuss challenges for the design of experimental paradigms for studying curiosity in psychology and cognitive neuroscience. Keywords: Curiosity, intrinsic motivation, lifelong learning, predictions, world model, rewards, free-energy principle, learning progress, machine learning, AI, developmental robotics, development, curriculum learning, self-organization.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (1)
  1. Pierre-Yves Oudeyer (95 papers)
Citations (58)

Summary

Computational Theories of Curiosity-Driven Learning

Pierre-Yves Oudeyer's paper, "Computational Theories of Curiosity-Driven Learning," explores the multifaceted nature and pivotal role of curiosity in facilitating learning processes in both biological organisms and artificial systems. This overview presents a detailed examination of curiosity-driven mechanisms, drawing parallels between human exploration and computational frameworks in robotics and AI.

Overview of Curiosity-Driven Learning

Curiosity plays a vital role in the organizational structure of learning, serving as an intrinsic motivator that enables organisms to navigate environments, solve complex tasks, and build robust world models. The paper delineates the significance of curiosity in situations where extrinsic rewards are sparse or deceptive. Curiosity-driven learning, by encouraging exploration beyond immediate rewards, fosters the discovery of diverse behavioral skills and facilitates the initial stages of learning without prior knowledge or guidance.

The paper provides an analysis of normative and heuristic computational models of curiosity. The normative models, illustrated by concepts such as the free-energy principle, predict optimal exploration strategies aimed at minimizing future surprises. In contrast, heuristic models, such as the learning progress hypothesis, emphasize mechanisms that maximize the rate of error reduction or learning progress. These frameworks together help elucidate the bidirectional causal links between curiosity and learning, proposing hypotheses about the role of curiosity in self-organizing developmental structures through curriculum learning.

Key Experiments and Theoretical Implications

Through developmental robotics experiments, the paper showcases the practical applications of curiosity-driven mechanisms outside of purely theoretical realms. The research demonstrates how robotic systems equipped with curiosity algorithms outperform traditional approaches in tasks characterized by rare rewards. By autonomously generating and prioritizing goals with high learning progress, artificial agents can navigate high-dimensional sensorimotor spaces, effectively self-organizing their developmental trajectories.

One notable experimental setup is the Playground Experiment, where a robot autonomously discovers a structured sequence of skills—from basic motor control to complex social interactions—guided solely by the intrinsic maximization of learning progress. This implicates curiosity as a central driver in the self-organization of developmental trajectories, directly mirroring observations in human infants.

Implications and Future Directions

The implications of Oudeyer's research are profound, offering insights into how curiosity mechanisms in humans and machines could be harnessed for enhanced learning and adaptation. While promising, the paper acknowledges the existing gap between theoretical models and experimental verifiability, encouraging the development of novel experimental paradigms in psychology and cognitive neuroscience that can rigorously capture and analyze curiosity-driven behaviors. The integration and testing of these mechanisms in evolving AI systems are projected to advance not only technology but also theoretical understanding of human cognitive processes.

Looking ahead, the paper suggests that curiosity-driven exploration should become a pivotal theme within cognitive and developmental sciences, inviting further inquiry into its broader applications and interactions with other motivational systems. As computational models continue to refine our understanding of curiosity, both practical implementations in AI technology and theoretical insights into human learning processes are expected to progress, creating a symbiotic growth in artificial intelligence and cognitive science.

Youtube Logo Streamline Icon: https://streamlinehq.com