MIT and Facebook released SLAC, a dataset of over 500,000 videos with nearly 2 million annotations across 200 activities. The team removed duplicates and created hard negatives, videos with the right context but missing the target action. Pretraining on SLAC boosted downstream task accuracy by up to 30 percent.
Dear fellow scholars, this is two Minute Papers with zsolnai. This project is about a dataset created.
Through a joint effort between MIT and Facebook.
As it turns out, this dataset is.
Way more useful than I initially thought. I'll tell you in a moment why datasets are used to train and test the quality of learning algorithms. This particular dataset contains short video clips.
These clips are passed to a neural network which is asked to classify the.
Kind of activity that is taking place in the video. In this dataset, there are many cases where everything is given to come to a logical answer that is wrong. We may be in a room with the climbing wall, but exercising is not necessarily happening. We could be around the swimming pool, but swimming is not necessarily happening. I'm pretty sure this has happened to you too. This is a brilliant idea because it is super easy for a neural network to assume that if there is a swimming pool, swimming is probably happening.
But it takes a great deal of.
Understanding to actually what constitutes the swimming part.
A few episodes ago we discussed that.
This could potentially be a stepping stone towards creating machines that think like humans. Without looking into it, it would be easy to think that creating a dataset is basically throwing a bunch of training samples together and calling it a day. I can assure you that this is not the case and that creating a data set like this was a Herculean effort as it contains more than half a million videos and almost 2 million annotations for 200 different activ activities and.
There are plenty of pre processing steps.
That one has to perform to make it usable. The collection procedure contains a video crawling step where a large number of videos are obtained from YouTube which are to be deduplicated, which means removing videos that are too similar to one already contained in the database. A classical case is many different kinds of commentary on the same footage. This amounted to the removal of more than 150,000 videos. Then all of these videos undergo a shot and person detection step where relevant subclips are extracted that contain some kind of human activity. These are then looked at by two different classifiers and depending on whether there was a consensus between the two, a decision is made whether the clip is to be discarded or not. This step helps balancing the ratio of videos where there is some sort of relevant action compared to the clips where there is no relevant action happening. This also makes the negative samples much harder because the context may be correct but the expected activity may not be there.
This is the classical hard case with.
The swimming pool and people in swimming suits twiddling their thumbs instead of swimming. And here comes the more interesting part when trying to train a neural network.
For other loosely related tasks, using this.
Dataset for pre training improves the score significantly. I'll try to give a little context for the numbers because these numbers are absolutely incredible.
There are cases where the success rate.
Is improved by over 30% which speaks for itself.
However there are other cases where the.
Difference is about 10 to 15%. That is also remarkable when we are talking about high numbers because the closer the classifier gets to 100% the more difficult the remaining corner cases are that improve the accuracy. In these cases even a 3% improvement is remarkable. And before we go, greetings and best.
Regards to Lucas the little scholar who.
Seems to be absorbing the papers along with the mother's milk. Excellent. You can start early enough in the pursuit of knowledge. Thanks for watching and for your generous support and I'll see you next time.
