Deep neural networks can be very effective for image classification, but they usually rely on regularization methods, transfer learning, and/or large supervised training sets to reduce/avoid high classification accuracy on the training data with low accuracy on unseen test sets --- i.e., a phenomenon known as data overfitting. Transfer learning is not always possible and large supervised training sets are usually impractical in applications that require experts for data supervision. Possible solutions include data augmentation and other strategies to improve the architecture and weights of the network from a limited number of supervised examples. We have investigated such solutions in the main project, FAPESP 2016/25776-0, through the use of Encoder-Decoder Neural Networks (EDNNs), Convolutional Neural Networks (CNNs), and Visual Analytics. We are interested in solutions that improve the design of a CNN for image classification by exploiting the EDNN and Visual Analytics methods to copy with the absence of supervised examples. Therefore, the purpose of this BEPE project is the design of an user interface for interactive machine learning based on EDNNs, CNNs, and Visual Analytics. Validation studies will be conducted with image datasets from distinct applications related to the thematic project, FAPESP 2014/12236-1, coordinated by the advisor.
News published in Agência FAPESP Newsletter about the scholarship: