Empirical Evaluation of Deep Convolutional Neural Networks as Feature Extractors
Alfred Kishek · Deep Blue (University of Michigan) · 2017
Convolutional Neural Networks (CNNs) trained through backpropagation are central to several, competition-winning visual systems. However, these networks often require a very large number of annotated samples, lengthy periods of training, and the estimation and tuning of a plethora of hyperparameters. This thesis evaluates the effectiveness of CNNs as feature extractors and assesses the transferability of their feature maps using a Support Vector Machine (SVM) for evaluation. To compare representations, the parameters learned from a CNN are transfered to various unseen datasets and tasks. The results reveal a significant performance gain on target tasks with small amounts of data. However, as the number of training samples increases, the performance advantage of using the extracted features is diminished and the resulting classifier has high variance.