UNSUPERVISED LEARNING OF VISUAL STRUCTURE USING PREDICTIVE GENERATIVE NETWORKS
Name
CBMM-Memo-040.pdf
Size
1.92 MB
Format
Adobe PDF
Checksum (MD5)
af15af571dcc6810f014cb5b6269978c
Author(s)
Lotter, William
Kreiman, Gabriel
Cox, David
Date Issued
December 15, 2015
Publisher
Center for Brains, Minds and Machines (CBMM), arXiv
Citation
arXiv:1511.06380v1
Series/Report no.
CBMM Memo Series;040
Abstract
The ability to predict future states of the environment is a central pillar of intelligence. At its core, effective prediction requires an internal model of the world and an understanding of the rules by which the world changes. Here, we explore the internal models developed by deep neural networks trained using a loss based on predicting future frames in synthetic video sequences, using an Encoder-Recurrent-Decoder framework (Fragkiadaki et al., 2015). We first show that this architecture can achieve excellent performance in visual sequence prediction tasks, including state-of-the-art performance in a standard “bouncing balls” dataset (Sutskever et al., 2009). We then train on clips of out-of-the-plane rotations of computer-generated faces, using both mean-squared error and a generative adversarial loss (Goodfellow et al., 2014), extending the latter to a recurrent, conditional setting. Despite being trained end-to-end to predict only pixel-level information, our Predictive Generative Networks learn a representation of the latent variables of the underlying generative process. Importantly, we find that this representation is naturally tolerant to object transformations, and generalizes well to new tasks, such as classification of static images. Similar models trained solely with a reconstruction loss fail to generalize as effectively. We argue that prediction can serve as a powerful unsupervised loss for learning rich internal representations of high-level object features.
Subjects
Neural Networks
Encoder-Recurrent-Decoder framework
Vision
Predictive Generative Networks
Neuroscience
Terms of Use
Attribution-NonCommercial 3.0 United States
Persistent DSpace Link
