Multi-function Convolutional Neural Networks for Improving Image\n Classification Performance

2018/05/29 by Luna M. Zhang, Zhang, Luna M.
Computer Science · Medicine · Neuroscience · #Advanced Neural Network Applications #Brain Tumor Detection and Classification #COVID-19 diagnosis using AI #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Neural Networks and Applications

paper · pdf · doi:10.48550/arxiv.1805.11788

openalex publication_date 2018/05/29 · openalex created_date 2022/10/01 · openalex updated_date 2026/07/28

Abstract

Traditional Convolutional Neural Networks (CNNs) typically use the same\nactivation function (usually ReLU) for all neurons with non-linear mapping\noperations. For example, the deep convolutional architecture Inception-v4 uses\nReLU. To improve the classification performance of traditional CNNs, a new\n"Multi-function Convolutional Neural Network" (MCNN) is created by using\ndifferent activation functions for different neurons. For n neurons and m\ndifferent activation functions, there are a total of mⁿ-m MCNNs and only m\ntraditional CNNs. Therefore, the best model is very likely to be chosen from\nMCNNs because there are mⁿ-2m more MCNNs than traditional CNNs. For\nperformance analysis, two different datasets for two applications (classifying\nhandwritten digits from the MNIST database and classifying brain MRI images\ninto one of the four stages of Alzheimer's disease (AD)) are used. For both\napplications, an activation function is randomly selected for each layer of a\nMCNN. For the AD diagnosis application, MCNNs using a newly created\nmulti-function Inception-v4 architecture are constructed. Overall, simulations\nshow that MCNNs can outperform traditional CNNs in terms of multi-class\nclassification accuracy for both applications. An important future research\nwork will be to efficiently select the best MCNN from mⁿ-m candidate MCNNs.\nCurrent CNN software only provides users with partial functionality of MCNNs\nsince different layers can use different activation functions but not\nindividual neurons in the same layer. Thus, modifying current CNN software\nsystems such as ResNets, DenseNets, and Dual Path Networks by using multiple\nactivation functions and developing more effective and faster MCNN software\nsystems and tools would be very useful to solve difficult practical image\nclassification problems.\n

Citations

• Gradient-based learning applied to document recognition
• Open Access Series of Imaging Studies (OASIS): Cross-sectional MRI Data in Young, Middle Aged, Nondemented, and Demented Older Adults

Related

• Effective, Fast, and Memory-Efficient Compressed Multi-function\n Convolutional Neural Networks for More Accurate Medical Image Classification
• A New Compensatory Genetic Algorithm-Based Method for Effective Compressed Multi-function Convolutional Neural Network Model Selection with Multi-Objective Optimization
• ASU-CNN: An Efficient Deep Architecture for Image Classification and Feature Visualizations
• Convolutional Neural Networks:
• Convolutional Neural Networks In Convolution