Learning from One Continuous Video Stream

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Carreira, João, King, Michael, Pătrăucean, Viorica, Gokay, Dilara, Ionescu, Cătălin, Yang, Yi, Zoran, Daniel, Heyward, Joseph, Doersch, Carl, Aytar, Yusuf, Damen, Dima, Zisserman, Andrew
Format: Preprint
Published: 2023
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914732723666944
author Carreira, João
King, Michael
Pătrăucean, Viorica
Gokay, Dilara
Ionescu, Cătălin
Yang, Yi
Zoran, Daniel
Heyward, Joseph
Doersch, Carl
Aytar, Yusuf
Damen, Dima
Zisserman, Andrew
author_facet Carreira, João
King, Michael
Pătrăucean, Viorica
Gokay, Dilara
Ionescu, Cătălin
Yang, Yi
Zoran, Daniel
Heyward, Joseph
Doersch, Carl
Aytar, Yusuf
Damen, Dima
Zisserman, Andrew
contents We introduce a framework for online learning from a single continuous video stream -- the way people and animals learn, without mini-batches, data augmentation or shuffling. This poses great challenges given the high correlation between consecutive video frames and there is very little prior work on it. Our framework allows us to do a first deep dive into the topic and includes a collection of streams and tasks composed from two existing video datasets, plus methodology for performance evaluation that considers both adaptation and generalization. We employ pixel-to-pixel modelling as a practical and flexible way to switch between pre-training and single-stream evaluation as well as between arbitrary tasks, without ever requiring changes to models and always using the same pixel loss. Equipped with this framework we obtained large single-stream learning gains from pre-training with a novel family of future prediction tasks, found that momentum hurts, and that the pace of weight updates matters. The combination of these insights leads to matching the performance of IID learning with batch size 1, when using the same architecture and without costly replay buffers.
format Preprint
id arxiv_https___arxiv_org_abs_2312_00598
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Learning from One Continuous Video Stream
Carreira, João
King, Michael
Pătrăucean, Viorica
Gokay, Dilara
Ionescu, Cătălin
Yang, Yi
Zoran, Daniel
Heyward, Joseph
Doersch, Carl
Aytar, Yusuf
Damen, Dima
Zisserman, Andrew
Computer Vision and Pattern Recognition
Artificial Intelligence
We introduce a framework for online learning from a single continuous video stream -- the way people and animals learn, without mini-batches, data augmentation or shuffling. This poses great challenges given the high correlation between consecutive video frames and there is very little prior work on it. Our framework allows us to do a first deep dive into the topic and includes a collection of streams and tasks composed from two existing video datasets, plus methodology for performance evaluation that considers both adaptation and generalization. We employ pixel-to-pixel modelling as a practical and flexible way to switch between pre-training and single-stream evaluation as well as between arbitrary tasks, without ever requiring changes to models and always using the same pixel loss. Equipped with this framework we obtained large single-stream learning gains from pre-training with a novel family of future prediction tasks, found that momentum hurts, and that the pace of weight updates matters. The combination of these insights leads to matching the performance of IID learning with batch size 1, when using the same architecture and without costly replay buffers.
title Learning from One Continuous Video Stream
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2312.00598