Dual Sum-Product Networks Autoencoder for Multi-Label Classification
- PDF / 374,104 Bytes
- 9 Pages / 612.284 x 810.709 pts Page_size
- 62 Downloads / 198 Views
Dual Sum-Product Networks Autoencoder for Multi-Label Classification WANG Shengsheng a (
),
ZHANG Hang b∗ (
),
í
CHEN Juan a (
)
(a. College of Computer Science and Technology; b. College of Software, Jilin University, Changchun 130012, China)
© Shanghai Jiao Tong University and Springer-Verlag GmbH Germany, part of Springer Nature 2020 Abstract: Sum-product networks (SPNs) are an expressive deep probabilistic architecture with solid theoretical foundations, which allows tractable and exact inference. SPNs always act as black-box inference machine in many artificial intelligence tasks. Due to their recursive definition, SPNs can also be naturally employed as hierarchical feature extractors. Recently, SPNs have been successfully employed as autoencoder framework in representation learning. However, SPNs autoencoder ignores the model structural duality and trains the models separately and independently. In this work, we propose a Dual-SPNs autoencoder which designs two SPNs autoencoders to compose as a dual form. This approach trains the models simultaneously, and explicitly exploits the structural duality between them to enhance the training process. Experimental results on several multilabel classification problems demonstrate that Dual-SPNs autoencoder is very competitive against with state-of-the-art autoencoder architectures. Key words: sum-product networks (SPNs), representation learning, dual learning, multi-label classification CLC number: TP 181 Document code: A
0 Introduction There is a remarkable interest in learning tractable representations facilitating exact probabilistic inference in polynomial time complexity[1] . Sum-product networks (SPNs)[2] were the first model that combines powerful representation and tractable inference ability. SPNs have been successfully applied in natural language processing[3] and computer vision[4] . So far, SPNs always act as a deep tractable probabilistic model that has mainly been used as a distribution estimator: applying some inference problem and predictive tasks. However, SPNs not only have exact inference ability, but also have powerful representation power. Vergari et al.[5] proposed an SPNs autoencoder model and used inner node embedding for representation learning. They proposed an encoder based on SPNs model to encode raw data into embedding representation and changed SPNs to max-product networks (MPNs) as a decoder model that decodes these embedding representations back into the original sample input space. The SPNs autoencoder can learn hierarchical probabilistic part-based features using an unsupervised Received date: 2018-06-01 Foundation item: the National Natural Science Foundation of China (No. 61472161), the Science & Technology Development Project of Jilin Province (Nos. 20180101334JC and 20160520099JH) ∗E-mail: [email protected]
method. The SPNs autoencoder also has several problems: There are structural duality in SPNs autoencoder which is largely ignored in current way; the feedback signal between SPNs encoder and MPNs decoder can not be
Data Loading...
