Papers of the day   All papers

Sideways: Depth-Parallel Training of Video Models

Comments

Jan 20 2020 DeepMind

Given the smoothness of videos, can we learn models more efficiently than with #backprop? We present Sideways - a step towards a high-throughput, approximate backprop that considers the one-way direction of time and pipelines forward and backward passes. https://arxiv.org/pdf/2001.06232.pdf https://t.co/evbwULE0s2
8 replies, 1015 likes


Jan 20 2020 Mateusz Malinowski

What if we drop the assumption that #backpropagation is instantaneous, and instead, we consider it as a process requiring time, and the time only moves forward. Can we use this property to pipeline forward and back passes, making it more parallel?@joaocarreira @GrzegorzMS Viorica
1 replies, 26 likes


Jan 20 2020 Andrew Davison

Looks like interesting work on computational patterns for efficient training from sequential video data. Efficient continual learning of persistent models from incremental (real-time) data is exactly what I think I've always been most interested in. #SpatialAI
1 replies, 24 likes


Jan 20 2020 Daisuke Okanohara

For training video models, we can use the activation computed from an adjacent frame to compute the gradient. This approximated gradient works well, and actually improves the generalization, which enables efficient depth parallelization. https://arxiv.org/abs/2001.06232
0 replies, 6 likes


Jan 20 2020 arXiv CS-CV

Sideways: Depth-Parallel Training of Video Models http://arxiv.org/abs/2001.06232
0 replies, 3 likes


Jan 26 2020 Alison B Lowndes ✿

Finally! Someone updates research papers. New #LaTeX, new backprop, from @DeepMind @MateuszOnAI, Viorica Pătrăucean, @GrzegorzMS @joaocarreira @sindero. PS double negative in the Conclusion hurting my brain?? https://arxiv.org/pdf/2001.06232.pdf
0 replies, 2 likes


Jan 31 2020 akira

https://arxiv.org/abs/2001.06232 Proposes an efficient learning method Sideaways for video analysis. Normally, BP is performed for each frame, but in Sideaway, parameters are updated using slightly future frame. Not only efficient, but also better accuracy because of regularization. https://t.co/hQBdJRUZgj
0 replies, 1 likes


Content