What can biological systems teach us about embodied learning?

Event: Workshop on Embodied Learning · Duration: 0 min · ▶ Watch on YouTube

Abstract

The presentation delves into the lessons biological systems, especially human children, can offer to the field of embodied learning in AI. It highlights the unique sensory and cognitive capabilities of biological agents, contrasting them with current AI training paradigms. The talk emphasizes that embodied interactions and developmental changes in sensory input significantly shape learning, leading to efficient few-shot learning despite weak supervision. By analyzing first-person egocentric data from children, researchers aim to understand how these natural learning curricula and data distributions could inspire more robust and generalizable machine learning models. Preliminary results suggest that child-like data distributions can lead to faster learning and improved generalization in computational models.

Speakers

  • David Crandall — Luddy School of Informatics, Computing, and Engineering, Indiana University
  • Chonghao Sima — Unknown
  • Chen Yu — UT Austin
  • Linda Smith — Indiana University
  • Justin Wood — Indiana University
  • Samantha Wood — Indiana University
  • Aditi Jayaraman — Indiana University
  • Sven Bambach — Indiana University
  • Satoshi Tsutsui — Indiana University
  • Zehua Zhang — Indiana University
  • Bardia Doosti — Indiana University
  • Shujon Naha — Indiana University
  • Majid Mirbagheri — Indiana University
  • Francisco Perelli — Indiana University
  • Michael Jones — Indiana University
  • Olaf Sooms — Indiana University
  • Richard Betzel — Indiana University
  • Zoran Tiganj — Indiana University
  • Amatuni — Unknown
  • Schroer — Unknown
  • Peters — Unknown
  • Reza — Unknown

Talks (1)

  • 00:00:06 — David Crandall: What can biological systems teach us about embodied learning?
    • This talk explores how studying biological learning systems, particularly human children, can provide insights for developing more efficient and robust embodied AI systems, focusing on the impact of embodied interactions and developmental curricula on learning.

Key Takeaways

  • Bodies drive learning by constraining our learning inputs, modalities, tasks, supervision, and objective functions, which are typically very different from those used in computer vision and across different animals.
  • We should actively seek opportunities to learn from the world’s best embodied learning system: the human child, to inform the development of more robust and efficient AI.
  • Better understanding how children solve learning tasks could be important for advancing developmental psychology and improving educational outcomes for kids.
  • Data collected by kids through natural, embodied interactions has a special structure (e.g., diverse views, changing distributions over time) that is efficient for learning, and computer vision may benefit from incorporating these insights.
  • Raw sensory data perceived by infant learners, including egocentric video, gaze, and parent utterances, contains enough statistical information for word learning by a simple ideal learning model.

Methods / Models / Datasets Mentioned

  • CNNs
  • Transformer models
  • VideoMAE
  • JEPA-TT
  • SimCLR-TT
  • ResNet
  • HomeView dataset
  • COCO dataset
  • ShapeNet dataset
  • Multidimensional scaling
  • Binary classifier

Topics

Embodied learning · Biological systems · Developmental psychology · Computer vision · Few-shot learning · Weakly-supervised learning · Egocentric vision · Child development · Machine learning curricula · Data distribution · Human-robot interaction

Notes

Open for commentary — connections to other work, critiques, follow-up reading.