ECCV 2018 Daily - Thursday

Daily Thursday 13 Octavia Camps is a professor of electrical and computer engineering at Northeastern University in Boston, MA. We speak to her about her team's poster yesterday on a dynamical atoms-based network for video prediction. Octavia tells us what the work is about: “ What we are doing here is predicting the next frame after seeing T-frames. So, see T and predict T+1, T+2, T+3. What we have done is incorporate a new type of layer for computer networks that encapsulates the dynamics of the input data. It’s a replacement of LSTM and designed based on first principles from system identification. ” The beauty of the work is that it allows you to build very compact models to capture dynamics . It is easy to train and easy to understand what the network is learning, because the parameters have a physical meaning. Octavia points out that if you know her past work, you will know she is a fan of dynamics and believes that dynamics first principles are good . She says this was hard to put into a neural network. Along with her team, she thought for a long time about how to bring their expertise in that area into deep networks. It took a couple of years to figure it out, but they finally got it, and in the end, it was a simple solution. 6 They finally got it, and in the end, it was a simple solution! DYAN - Octavia Camps Poster Presentation