Thu Jun 30 2022
Wed Jun 29 2022

Beyond neural scaling laws: beating power law scaling via data pruning

Neural Networks
Deep Learning
Data Pruning
CIFAR-10
SVHN
ImageNet

The paper suggests that the discovery of good data-pruning metrics may provide a viable path forward to substantially improved neural scaling laws.

The improved scaling prediction can reduce the resource costs of modern deep learning.

https://arxiv.org/pdf/2206.14486.pdf
https://arxiv.org/abs/2206.14486
https://twitter.com/arankomatsuzaki/status/1542314522454552576/photo/1

Masked World Models for Visual Control

Robotics
Visual Model-Based Reinforcement Learning
Machine Learning
Meta-world
RLBench

The paper achieves SotA performance on various visual robotic tasks from Meta-world and RLBench using an approach inspired by MAE.

The decoupling approach achieves state-of-the-art performance on a variety of visual robotic tasks.

https://sites.google.com/corp/view/mwm-rl
https://arxiv.org/pdf/2206.14244.pdf
https://arxiv.org/abs/2206.14244
https://twitter.com/arankomatsuzaki/status/1542309982682501120/photo/1
Tue Jun 28 2022
Mon Jun 27 2022
Sun Jun 26 2022
Thu Jun 23 2022