Toss a ball to a human and they’ll find a multitude of ways to interact with it — kicking, dribbling, heading or even deflating the thing. It’s not as easy as one might imagine to train an AI model to accurately predict what a human will do next, even when they are interacting with a relatively simple object like a ball.
When such next-state uncertainty presents in a video, the task of generating future frames given context frames — known as video prediction — is notoriously difficult. Along with the prediction process, there are also many spatio-temporal variables to deal with in video generation.
Most existing methods include handcrafted measures in the neural networks, for example to separate information streams, add high-level information like landmarks and semantic segmentation masks, or perform specialized computations like warping, optical flow, background masking, etc. Other methods have been applied in relatively simpler environments such as scenes that mainly involve synthetic shapes or human faces or bodies.
New research from the University of Michigan, Google, and Adobe Research questions whether such handcrafted architectures are necessary. In what they say is the first large-scale study of the effects of minimal inductive bias and maximal capacity on video prediction, the researchers show that it’s possible to generate high quality video prediction simply by increasing the scale of computation.
The researchers trained large models on three different datasets — one for modeling object interactions, another for modeling human motion, and a third for modeling car driving.
In their experiments they found that maximizing the capacity of a standard neural network can achieve higher quality video predictions. Recurrent models tend to outperform non-recurrent models, with large-scale recurrent neural networks (RNNs) achieving SOTA performance in predicting what’s coming next in videos. The researchers also noted that stochastic models perform better than non-stochastic models.
The authors are calling for further studies to help discover an ideal combination of minimal inductive bias and maximal model capacity for optimizing video prediction.
The paper High Fidelity Video Prediction with Large Stochastic Recurrent Neural Networks is on arXiv.
Journalist: Yuan Yuan | Editor: Michael Sarazen