Repository logo
 

One-shot learning with pretrained convolutional neural network

Date

2019

Authors

Yu, Zhixian, author
Draper, Bruce, advisor
Beveridge, Ross, committee member
Peterson, Chris, committee member

Journal Title

Journal ISSN

Volume Title

Abstract

Recent progress in convolutional neural networks and deep learning has revolutionized the image classification field, and computers can now classify images with a very high accuracy. However, unlike the human vision system which efficiently recognizes a new object after seeing a similar one, recognizing new classes of images requires a time- and resource-consuming process of retraining a neural network due to several restrictions. Since a pretrained neural network has seen a large amount of training data, it may be generalized to effectively and efficiently recognize new classes considering it may extract patterns from training images. This inspires some research in one-shot learning, which is the process of learning to classify a novel class through one training image from the novel class. One-shot learning can help expand the use of a trained convolutional neural network without costly model retraining. In addition to the practical application of one-shot learning, it is also important to understand how a convolutional neural network supports one-shot learning. More specifically, how does the feature space structure to support one-shot learning? This can potentially help us better understand the mechanisms of convolutional neural networks. This thesis proposes an approximate nearest neighbor-based method for one-shot learning. This method makes use of the features produced by a pretrained convolutional neural network and builds a proximity forest to classify new classes. The algorithm is tested in two datasets with different scales and achieves reasonable high classification accuracy in both datasets. Furthermore, this thesis tries to understand the feature space to explain the success of our proposed method. A novel tool generalized curvature analysis is used to probe the feature space structure of the convolutional neural network. The results show that the feature space curves around samples with both known classes and unknown in-domain classes, but not around transition samples between classes or out-of-domain samples. In addition, the low curvature of out-of-domain samples is correlated with the inability of a pretrained convolutional neural network to classify out-of-domain classes, indicating that a pretrained model cannot generate useful feature representations for out-of-domain samples. In summary, this thesis proposes a new method for one-shot learning, and provides insight into understanding the feature space of convolutional neural networks.

Description

Rights Access

Subject

convolutional neural network
image recognition
proximity forest
generalized curvature analysis
approximate nearest neighbor
one-shot learning

Citation

Associated Publications