A neural network learns to synthesize human and animal motion from messy, unlabeled data. Dataset augmentation multiplies training data 10 to 30 times, letting the system generate diverse behaviors like dribbling, hopping, and punching with interactive control.
October 19, 2019.
Up to 10 to 30 times more training data.
Can train ImageNet to 93% accuracy for less than $19.
This episode has been supported by Lamda. Dear fellow scholars, this is two minute papers with zsolnai. Last year, an amazing neural network based technique appeared that was able to look at a bunch of unlabeled motion data and learned to weave them together to control the motion of quadrupeds. Like this wolf here, it was able to successfully address the shortcomings of previous works. For instance, the weird sliding motions have been eliminated and it was also capable of following some predefined trajectories. This new paper continues research in this direction by proposing a technique that is also capable of interacting with its environment or other characters. For instance, they can punch each other and after the punch they can recover from undesirable positions and more. The problem formulation is as it is given the current state of the character and the goal, and you see here with blue how it predicts the motion. To continue, it understands that we have to walk towards the goal, that we are likely to fall when hit by a ball, and it knows that then we have to get up and continue our journey and eventually reach our goal. Some amazing life advice from the AI right there. The goal here is also to learn something meaningful from lots of barely labeled human motion data. Barely labeled means that a bunch of videos are given almost as is, without additional information on what movements are being performed in these videos. If we had labels for all this data that you see here, it would say that this sequence shows a jump and these ones are for running. However, the labeling process takes a ton of time and effort, so if we can get away without it, that's glorious. But it in return with this, we create an additional burden that the learning algorithm has to shoulder. Unfortunately, the problem gets even worse. As you see here, the number of frames contained in the original dataset is very scarce. To alleviate this, the authors decided to augment this dataset, which means trying to combine parts of this data to squeeze out as much information as possible. You see some examples here how this motion data can be combined from many small segments. And in the paper they show that the augmentation helps us create even up to 10 to 30 times more training data for the neural networks. As a result of this augmented dataset, it can learn to perform zombie gorilla movements, chicken hopping, even dribbling with a basketball, you name it. What's even more, we can give the AI high level commands interactively and it will try to weave the motions together appropriately. They can also punch each other. Ow. And all this was learned from a bunch of unorganized data. What a time to be alive. If you are a researcher or a startup looking for cheap GPU compute to run these algorithms, check out Lambda GPU Cloud. I've talked about Lambda's GPU workstations in other videos and am happy to tell you that they are offering GPU Cloud services as well. The Lambda GPU Cloud can train ImageNet to 93% accuracy for less than $19. Lambda's web based IDE lets you easily access your instance right in your browser and finally, hold on to your papers because the Lambda GPU Cloud costs less than half of AWS in and Azure. Make sure to go to lambdalabs.compapers and sign up for one of their amazing GPU instances today. Thanks for watching and for your generous support and I'll see you next time.
