Prednet---Deep predictive coding networks for video prediction and unsupervised learning ICLR 2017
2017.03.12
Code and video examples can found at: https://coxlab.github.io/prednet/
Absrtact: Deep learning techniques based on supervised training have achieved great success, but unsupervised problems remain a difficult problem (learning from the data that has never been labeled a domain structure). This article explores the problems of video prediction in unsupervised learning. A kind of "prednet" structure was designed, the project was realized, and the experimental results were very gratifying. The experimental results show that the prediction represents a very powerful unsupervised learning framework that can potentially learn from objects or scene structures.
Network design:
As shown in the process, there is a series of modules stacked together to produce. The network first makes local predictions and then subtracts the actual input into the next layer.
Simply put, each module can be divided into 4 parts:
1. An input convolutional layer $A _l$
2. Circular presentation layer $R _l$
3. Prediction Layer $\hat{a}_l $
4. Error presentation layer $E _l$
$R _l$ is a cyclic convolution network that produces a predictive a^l,layer input that is Al. The network calculates the difference between Al and a^l, then outputs an error representation of El, divided into individually modified positive and negative error passes. The error El is passed to the convolution layer as the input $A _{l+1}$ of the next layer. $R _l$ module has two inputs from: directly copied El, and the next layer $R _{l+1}$ input.
This network can be divided into two most important parts, the left Rl part is a cyclic production deconvolution network; Right Al and El are standard deep convolutional networks.
The goal of the model training is: Minimize the weighted sum of the firing rates of the error units. The error units here is similar to L1 error. Although there is no attempt here, you can also try other loss function.
The overall algorithm framework is as follows:
Experimental section:
Prednet---Deep predictive coding networks for video prediction and unsupervised learning---paper notes