Blog posts

Sorted by categories or tags

2019

Embodied Multimodal Multitask Learning

2 minute read

Published:

Summary

  • Deep RL has been successful in training visual navigational agents conditioned on language for multimodal tasks, e.g. semantic goal navigation and embodied question answering
  • Propose a multitask model that can jointly learn these tasks
  • Novel Dual-Attention unit to disentangle words and visual concepts
  • Paper link

Toddler-Inspired Visual Object Learning

3 minute read

Published:

Summary

  • Real-world learning systems are limited in the quality and quantity of training datasets they can collect
  • Use head-mounted cameras and gaze trackers to collect egocentric images from human child in naturalistic learning contexts
  • Child data produces better object models than egocentric adult data
  • Child data exhibits unique combination of quality and diversity
  • Paper link

Automated Curriculum Learning for Neural Networks

3 minute read

Published:

Summary

  • Maximize learning efficiency by following a curriculum
  • Measure of the amount that the network learns from each data sample used as reward
  • Nonstationary multi-armed bandit algorithm
  • Consider a variety of signals based on rate of increase in prediction accuracy and network complexity
  • Experimental results with LSTMs on three curricula
  • Paper link