We proposes an automatic composition system using images. This system generates a monophonic melody from an input image in a symbolic domain. This method consists of 3 models including Composer, Evaluator, and Melody Generator Models. Composer is a Convolutional Neural Network (CNN) based model, which creates melody from an image, is trained with Evaluator by adversarial learning. Evaluator is a Long-Short Term Memory (LSTM) based model to evaluate phrases generated by Composer. Melody Generator is an LSTM based model to generate melody from phrase. We created three types of datasets including a new image-phrase dataset to train these models respectively.
Zhang et al. (2015) provided promising results of Character-level convolutional approaches to text understanding on Machine Translation and Text classification tasks.
The goal in this research was to apply such approaches to Twitter data, in order to see how well they are able to cope with raw social data.
The task was to predict hashtags of tweets based on their content only. Therefore, models build distributed representations of tweets using ConvNets and Gated Recurrent Units then rank hashtags by relevance to the tweet representation.
Results have shown to be on par with tweet2vec model by Dhingra et al. (2016), while being faster up to threefold during training on GPU.
References:Zhang et al. (2015) : https://arxiv.org/abs/1509.01626 Dhingra et al. (2016) : https://arxiv.org/abs/1605.03481