Discipline: Mathematics and Statistics
Subcategory: Mathematics and Statistics
Session: 1
Room: Tyler
Zackary A Kinsman - Texas Southern Univerity
Co-Author(s): Zackary Kinsman, Dr. N. Kulathunga, Dr. Y. Wang, Dr. D. Vrinceanu, Texas Southern University, Houston, Texas
The objectives of this research were to classify food images using deep convolutional neural networks (DCNNs) and to quantify how the non-linearity of the activation function affects the predictive power of the network. By using features from the pre-trained model Inception-V3 to reduce the dimension of the feature map and by attaching these features to a convolutional neural network (CNN) with a dense layer neurons, TensorFlow, and Keras, we developed convolutional neural network that classified images from the Food-11 data-set (École Polytechnique Fédérale de Lausanne, Multimedia Signal Processing Group) with an accuracy of over 79%. By investigating accuracy as a function of different alpha values in the activation function Leaky Rectified Linear Unit (Leaky ReLU), where alpha is a quantity that describes the linearity of the activation function, results were found about how the non-linearity in the neural network can affect the model’s overall performance. We discuss the results from different model architectures with different number of nodes and hidden layers and the possible causes for under-fitting and over-fitting. The strength of the non-linearity in the model can be adjusted by changing the negative slope in the Leaky ReLU activation function. We studied the trends in model accuracy as a function of the strength of non-linearity by adjusting the negative slope alpha in Leaky ReLU and found that the accuracy decreases as alpha increases.
Funder Acknowledgement(s): National Science Foundation
Faculty Advisor: Dr. Yunjiao Wang, zackary.kinsman@tsu.edu
Role: I was involved with each step of the process. I developed the model architecture of the neural network we used, tested the model using different loss and optimization functions, created the learning curves used to determine the best loss and optimization function combination, and created a new model that is currently being used to investigate how non-linearity in a CNN can affect the model's accuracy.