You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
A major benefit of self-supervision is we can truly scale and adapt on the fly. It could be 10% behind supervised ImageNet, it would still do better in real life. We show in http://online-objects.github.io that the longer our model looks at objects, the better it understands them.pic.twitter.com/fqi3hJCYIk
Self-supervision allows you to train on your test data, so it’s pretty much guaranteed to do better than a supervised model trained offline. In this video our model converges to ~2% identification error after 160s while the offline baseline trained on ImageNet is stuck at ~50%.pic.twitter.com/YEDBB4ir2m
Our robot collects its own data, trains itself on it with our self-supervised objective using contrastive learning, then it can point to similar never-seen-before objects to the one in front of it, demonstrating generalization of object attributes.pic.twitter.com/ufOMpjawcb
Our model recovers object attributes, colors, shapes and classes entirely from scratch, without any labels, as shown in the nearest neighbors here (ordered left to right by embedding distance to the leftmost object).pic.twitter.com/S4IDAins2C
One caveat is that we don’t actually train online (yet), we just show what’s possible if you did. But training on the fly is not even necessary in the case of a robot deployed in a new home, it can spend its first few days looking around, train itself overnight and overfit to it.
We present this paper today at the LUV and DeepVision workshops at #CVPR2019.
Paper: http://online-objects.github.io
Authors: @_pirk_, Mohi Khansari, Yunfei Bai, @coreylynch, @psermanet
how robust is self-supervision, can it learn too much until catastrophic forgetting?
Do you have the code?
Do you see any benefit in that sort of framework where the model of the world has been trained on these random plays and therefore has good coverage of possibilities. And then training an model-based RL agent on some expert trajectories for specific tasks downstream.
@psermanet very interesting talk! Would you say that this idea of self-play can be used in a model-based RL framework to build up the dynamics of the world model? As I understood from your talk no RL was applied in any of these 3 methods of TCN, OCN and TCC.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
