Arindam Paul

Deep Mining Pinterest

Large, deep convolutional neural networks (based on both a self-developed library on theano and torch7) are trained to classify ~1 million real-life images from Pinterest, into 1300+ different classes. Architectures including AlexNet, VGG, GoogLeNet are experimented. We employ frequent itemset mining to group labels into superclasses, and embed this information into the loss function to improve accuracy.