Abstract: In this paper, we study the factors which determine the capacity of a Convolutional Neural Network (CNN) model and propose the ways to evaluate and adjust the capacity of a CNN model for best matching to a specific pattern recognition task. Firstly, a scheme is proposed to adjust the number of independent functional units within a CNN model to make it be better fitted to a task. Secondly, the number of independent functional units in the capsule network is adjusted to fit it to the training dataset. Thirdly, a method based on Bayesian GAN is proposed to enrich the variances in the current dataset to increase its complexity. Experimental results on the PASCAL VOC 2010 Person Part dataset and the MNIST dataset show that, in both conventional CNN models and capsule networks, the number of independent functional units is an important factor that determines the capacity of a network model. By adjusting the number of functional units, the capacity of a model can better match the complexity of a dataset.
Abstract: Obtaining labeled data in supervised learning is often
difficult and expensive, and thus the trained learning algorithm tends
to be overfitting due to small number of training data. As a result,
some researchers have focused on using unlabeled data which may
not necessary to follow the same generative distribution as the labeled
data to construct a high-level feature for improving performance on
supervised learning tasks. In this paper, we investigate the impact of
the relationship between unlabeled and labeled data for classification
performance. Specifically, we will apply difference unlabeled data
which have different degrees of relation to the labeled data for
handwritten digit classification task based on MNIST dataset. Our
experimental results show that the higher the degree of relation
between unlabeled and labeled data, the better the classification
performance. Although the unlabeled data that is completely from
different generative distribution to the labeled data provides the lowest
classification performance, we still achieve high classification performance.
This leads to expanding the applicability of the supervised
learning algorithms using unsupervised learning.