Online Convolutional Neural Network for Image Streams Classification
Tianxiang Zheng, Zhijie Wen · 2022
In recent years, convolutional neural networks (CNN) have been rapidly developed for offline learning and have shown powerful capabilities in areas such as image classification. However, since these models require sufficient time and batch data learning convergence in offline learning mode, this is not scalable for many realistic scenarios where new data arrives sequentially in streams. In response, we propose a new model: Online Convolutional Neural Network (OCNN). Unlike traditional CNN model, the OCNN is very flexible where its structure can change dynamically as data streams come in, allowing it to work well online. In particular, we use CNN as the feature extractor of the model, and a fully connected neural network (DNN) is connected after each feature extractor as the classifier. By using Hedge Backpropagation method to assign weights to each classifier and get the prediction results, OCNN can dynamically learn to adaptive depth from a sequence of data streams. We validated the efficacy of our model on several large-scale datasets: we improved the classification accuracy by 4.95 and 20.74 compared to baseline model on Fashion-MNIST and CIFAR-10 data streams.