How Evolution May Work Through Curiosity-Driven Developmental Process

Summary

• Discusses similarities between curiosity-driven learning trajectories in robots and developmental trajectories in infants
• Argues how these developmental structures can arise from evolution
• Links: [ website ] [ pdf ]

Background

• Infants generally do not learn passively and are seemingly motivated towards “interesting” sensorimotor activities
• Curiosity-driven learning can also be thought of as reducing uncertainty
• Curiosity is only one of several motivational mechanisms operating within organisms, many motivations interact at any given time

Methods

• Robotic experimental setup (see Oudeyer & Kaplan 2016 for more details):
  • Quadruped robot on infant play mat with “biteable” elephant, “bashable” hanging toy, and vocally imitating “teacher” robot
  • Motor primitives:
    • turning the head
    • opening/closing the mouth with varying strengths and timing
    • rocking the leg with varying angles and speed
    • vocalizing with varying pitches and lengths
  • Sensory primitives:
    • detect visual movement
    • salient visual properties,
    • proprioceptive touch in mouth
    • pitch and length of perceived sounds
  • Select action to maximize learning rate on predicting consequences of actions given sensory contexts

Results

• Refer to Oudeyer & Kaplan 2016 for detailed results and analysis
• In general, this learning process results in a developmental trajectory with a mixture of regularities and diversities given similar intrinsic motivation mechanisms and environments
• Self-organization (“universal” developmental phases):
  • Unorganized body babbling
  • Isolated random exploration of motor primitives
  • Directed non-affordant exploration towards objects
  • Affordance-based exploration of objects
• However, even with the same mechanisms and parameters, individual trajectories may invert stages within a phase accounting for emergent differences among individuals

Conclusion

• Results with vocalization suggests that language can be partially explained by general curiosity-driven exploration in addition to language-specific motivation
• Curiosity-driven development is in between learned “tabula rasa” and innate “programs”, arising from the dynamic interaction between cognitive mechanisms, properties of the body, and the pysical and social environment
• Evolutionary origins of intrinsic motivation can be explained by maximizing long-term fitness in rapidly changing environments, which rewards learning over domain-specific adaptations
• Structures learned through curiosity can later be recruited for unforeseen functions
• Weighting of motivational mechanisms (e.g. curiosity, hunger, mating) may account for observed species-typical development
  • Understanding how curiosity competes and complements these other forms of motivation is an interesting research direction