Abstract:
Convolutional Neural Networks (CNNs) showed state-of-the-art accuracyin image classification on large-scale image datasets. However, CNNs showconsiderable poor performance in classifying tiny data since their large number ofparameters over-fit the training data. We investigate the classification characteristicsof CNNs on tiny data, which are important for many practical applications. Thisstudy analyzes the performance of CNNs for direct and transfer learning-basedtraining approaches. Evaluation is performed on two publicly available benchmarkdatasets. Our study shows the accuracy change when altering the DCNN depth indirect training to indicate the optimal depth for direct training. Further, fine-tuningsource and target network with lower learning rate gives higher accuracy for tinyimage classification.
